I would be really curious to hear from devs at Databricks what the experience of development is like internally. I work at a small startup with essentially unlimited AI spend budget - the entire point is that I should be turning to it at every opportunity since our human labor is so expensive relative to tokens. So generally it's like:
- Spend most time prioritizing/discussing what to do.
- Once that's agreed, use Fable 5 High + 5.6 Sol XHigh come up with a design + plan. Agree on the high level plan. (Usually this just comes down to choosing where the change belongs on the spectrum between minimal patch <-> full redesign)
- Use Opus 5 or Sol Med to execute
- Auto-fix bugs and CI until green + thermonuclear review skill x3.
- Manual interrogation of change/nits
- Come up with QA plan and have Codex Computer Use execute on it
- Manually spot check the final result (usually a sizable diff, thousands of lines, complete feature E2E, etc)
I probably spend like $80 a day at least but I produce the output of 3 or 4 2022 engineers and probably at better quality. So it's easily worth it. Would I save money by switching to GLM 5.2 and such...perhaps? IDK. At our scale it's not worth the time spent building the eval harness to actually understand the performance tradeoff.
In our team's experience, the product of agents is generally The Homer (1). It does work, but it's vastly overengineered.
When I personally want tight code, I have to spend a considerable amount of time adjusting it manually:
- It needs to be trimmed down. In my experience, at least one agent I use struggles to produce minimalist designs, and it's very frustrating
- I need to consider whether there are solutions based on higher-level assumptions, that AIs typically miss
- I need to check whether there are off-the-shelf solutions - AIs like to reinvent the wheel
IMO, software production has become a mass-produced commodity in every sense - it's much more expensive to produce software manually, but the quality is not the same.
It took me a few seconds of deliberating if The Homer was a reference to baseball or "The Odyssey" and then realized there was a footnote
> software production has become a mass-produced commodity in every sense - it's much more expensive to produce software manually, but the quality is not the same.
This was more true a few months ago but Fable has improved the situation considerably.
Also just remember - minimalist code looks and feels great but customers do not read your code. I have caught myself many times providing "corrections" to abstractions that were already ~fine, just not perfect. The average SWE costs $200/hr. Careful you don't burn $50 worrying about code that will likely be rewritten or can be better abstracted when that's actually needed.
This is a pointless quibble but the hourly rate claim is not true--it's like ~$60 in the USA [0]. Maybe you meant at a specific Org but this is important context when comparing "pricing" between human and AI.
How is this not true? Taking a Senior SWE @ ~$200K, even just the base salary cost / 2080 working hours is $100/hr. Fully loaded employer cost + accounting for non-coding time gets you to upper 100s easily.
Even for a junior making $100K, I have a hard time believe their time is worth less than $75/hr or so.
Edit: Fine, "Senior" is not "Average". But naive salary is not the true numerator.
An MBA's rule of thumb is that a full time employee's hourly cost to a business is at least 1.5x to 2x times their salary depending on employer taxes, benefits, offices, travel, training, hardware, perks, etc.
"Will I benefit from this code being minimalist before [date]", where [date] is whenever you think the agent will be good enough to come back and make the corrections you would make today.
Minimalist code is necessary to keep AI agents working well for longer than a month on a system IME. At a certain point, their own machinations overwhelm them and they both slow down, and make worse and worse decisions.
No surprise, LLM companies optimize for waste. More tokens, and more prompts means more revenue. Reminds of Google’s Prabhakar Raghavan story: deliberately making search worse [1]
IME this works until it does not. This approach works well at the beginning of a greenfield project, but at the same time because it is so easy to add features, you will likely ship something that is way too over engineered. And that complexity will not amortize over next increments and will more likely lead to the entire project being a black box only fully understood by AI. However a more careful use of AI for targeted surgical changes is far more ”productive” in the long term IMO.
Something isn't clear about the size of your codebase here and the level of reliability your customers expect, as a reader of your comments. Clarity there will help.
My observation has been:
- Initial greenfield work by an LLM is fast and very effective with minimal or no human oversight.
- Subsequent work ends up being over engineered and very verbose. Assumptions are made that aren't suited to the problem at hand (for example I find Fable is extremely regex happy where structured data would work much better from a readability perspective.)
- Once code bloats beyond a certain point due to unguided LLM usage, complexity is high enough that only LLMs can operate on the codebase with any economical amount of time.
- Rinse repeat and your code ends up unclear about any state that's not explicitly being tested and verified in QA loops
For some of our products this has been fine, for others it's been problematic. An understanding of your size and reliability requirements will help make the conversation more productive.
Why aren't you guiding your LLM usage? Is that what I said - to spam it and not guide anything? Or to have a careful workflow where you agree on design and maximize your human judgement/leverage?
> any state that's not explicitly being tested and verified in QA loops
As opposed to before, when engineers perfectly reasoned about code behavior from first principals and QA was unnecessary?
> Why aren't you guiding your LLM usage? Is that what I said - to spam it and not guide anything? Or to have a careful workflow where you agree on design and maximize your human judgement/leverage?
You didn't say anything positively or negatively regarding this so I made an assumption that you were using the LLM relatively unguided (e.g. a bit of oversight, not the kind of thing that heavy code reviews used to involve pre-agents.) Feel free to add clarity on your actual usage loop.
> As opposed to before, when engineers perfectly reasoned about code behavior from first principals and QA was unnecessary?
In my experience, most engineers are quite good at reasoning about code behavior for non-QAed code paths. Obviously things fall through the cracks. But I've been in the ground floor of plenty of Big Techs in their early stages before agents and, yes, a lot of initial development had spotty test coverage and yet most of the engineers had good mental models of what was happening. It used to be a very valuable skill to wrap your head around a torrid piece of code with few or no tests but was nonetheless a core piece of your application. Conversely, agentic development can bring cognitive debt [1].
===
This isn't a fight. We aren't sparring over what's right and wrong. I'm just curious how other people use agents in their work as someone who is also now in a startup that uses LLM agents heavily and has no limitations on spend.
What is the point of working at a startup if you’re dealing with millions of lines of legacy code ? Isn’t the whole point of startups to create & innovate with a clean slate and modern tools?
I don't know if that's the whole point, but I agree with the sentiment, why would a startup be working in legacy code and where would that code come from if this is truly the start of something.
OP might just be working at a small software company or for one that broke from a bigger one and is now "startup" like?
"startup" and "legacy codebase" are diametrically opposed concepts.
And if you're saying (based on your other comments) that a 6 month window is enough to create a legacy codebase...that indicates a serious lack of experience or understanding as to what a legacy codebase is, or why they exist.
exactly. Usually legacy code forms when people lose context and confidence in parts of the codebase due to staff turnover etc and ppl avoid touching or enhancing those parts for long periods. Six months is a short time to accrue that much tech debt, its enough time where most of the people who created that "legacy" are probably still around. As you said indicates bigger problems.
Do you have issues with performance at the moment? Right now I tend to find that it produces absolutely terrible design patterns and especially performance. I mean maybe I don't know exactly what area you're looking at but yeah for us we tend to find it's terrible wrt dB/caching/scaling and often any performance improvements it proposes end up actually shooting itself in the foot and being worse than before but it's not very good at testing in an organized way to even notice it made it worse despite repeated prompts to do so I mean if I prompt it to test performance in a handheld structured way (it is very bad at finding out what performance to test and why) before making changes I can usually figure it out but it usually takes insistence on the specifics to really ensure a good solution that will actually fix the problem
Are you using the SOTA models at very high reasoning during planning? IME that makes a LOT of a difference. I‘d also never let them just rip into the architecture, but always push back and ask for alternatives first. Once the overall plan is nailed, not that much can go wrong. Provided it’s a reasonable change set and not a 20k LOC PR.
Performance is better than ever. It's never been more practical to set up wildly complex synthetic test environments and measure perf wins. Plus the models will find every possible algorithmic/design improvement.
It actually gives me quite an uncanny feeling, bulldozing over years of human optimization work with a newer, "perfect" design. Like bringing an AK-47 back to the middle ages.
Yes 100%. This morning I casually prompted Codex to drive the browser to complete extensive performance testing in-situ that would have literally been weeks of work before. Probably in reality it just wouldn't have been done, and performance guarantees would have been attempted up front via more careful design.
In this case the design was also AI generated, and there were limited wins to be found because the design was already superb.
Are you at least conversational in the subject matter? You're gonna have a good time just by paying attention and adjusting your workflow. If you're getting a lot of back and forth with it, its asking a lot of planning type questions, stop, step back, rethink the whole feature, and start again from the beginning with everything more fleshed out.
If you are in a brand new field, there's no way to bridge that divide. The issue is you don't know what is good or bad, or whether what you have learned is good or bad. You're in a sports car and you don't know how to drive much less what's track and what's field.
You can spend a lot of effort getting good at prompting towards writing tests and E2E tests to at least verify your app does what you expect it to, regardless of experience.
This is a great point and I agree. My own productivity varies based on what part of the codebase I'm working on. If it's "been in there before" and I know the right questions to ask, I can one-shot a good design/improvement. If I'm spending 20-30 minutes asking Fable to "draw a diagram so I can understand" - probably less so. But notably, I CAN get there in a fraction of the time it would have taken before. You can general personalized onboarding docs to ~anything.
Keep the decision-making and execution separate. Use the high IQ models to chat about the design and make them drive subagents to do the actual work. "Chat" style threads are actually quite cheap. Where it gets expensive is having Fable 5 output thousands of lines of implementation where 95% of it was already overdetermined and there were only a few important judgement calls.
I actually have no doubt that I could replace my Opus 5 Low/Medium subagent profiles with Grok 4.5/GLM 5.2/Deepseek v4 Flash and perf would probably be pretty similar.
On top of that - highly recommend adding accurate cost counters to your statusline. You can't improve what you don't measure! (Or even have any intuition about).
It's a fair point, it's not truly unlimited and I do wonder how that would change my workflow. I can definitely imagine if I was inside Anthropic or OAI with unlimited "fast" tokens, you would be more tempted to hand over even more of this process. I completely understand why they talk about "graph engineering" and such, my entire workflow above could be a graph and I could try to increase my leverage even further. Realistically though I am bounded by product decision making, not code output right now.
Kudos to databricks, I also find it interesting that such different companies (Stripe, Ramp, Databricks) are all building the exact same internal tools.
I think building companies is going to look more generic in the future because intelligence is an API now.
There are a surprising number of articles like this along the lines of, "we started using AI tools and ended up spending millions per year".
On what planet do people start paying for things without keeping an eye on the costs and no-one notices until you have spent a crazy amount? I don't understand. You are either paying a fixed amount which you are happy about in-advance or you are PAYG in which case you would ballpark how much it costs.
Otherwise it reads a bit like a fake problem, because it didn't really happen, you just foresaw it (as you should) and added a few guide rails.
The issue is the growth rates can cause costs to drastically change quickly. If you have 1,000 employees and the average is spending $100/month you're at a $1.2M run rate. But suddenly a new model comes out that's twice as expensive, there are some changes to the harness (we found randomly Claude Code and other harnesses will make changes that drastically impact efficiency), and then maybe you have some organic user growth as well and BOOM suddenly you're at a $10M run rate within 60 days. And it's now impossible to forecast future growth.
It is true that this problem can be mostly managed by the techniques we mention here. Those are actually pretty difficult to set up at scale, so many companies (including us) we only really did this in earnest once we started to see those large cost oscillations.
The main reason we shared this here is to maybe help other companies get infrastructure in place before massive cost swings rather than after.
Something underlying a lot of this is that pricing models for enterprise coding tools have changed from seat-based to consumption-based pretty quickly, as AI usage has exploded. For months, engineers were able to use unlimited AI for no marginal cost, but that's changed quickly.
In addition, we're seeing people applying AI to more and more use cases, so token growth is very significant. Paired with consumption pricing, it's brought this problem to the forefront very quickly for lots of companies.
Not only this, but perhaps even more nefarious is that AWS gives lots of startups $100k+ in credits. This feels generous when you get it. In reality, it means that (unless you are in a compute intensive startup) you can go for months or years before you hit this, but by the time you do, you already have very solid monthly spend.
Initially, you picked the Multi-ZA RDS db.t3.2xlarge instance because you figured "eh i have credits anyway". Two years later, someone looks at this and says "hey, this is expensive and I bet we can do everything we need on a machine half the size". But then they think "if i downsize it and that works, i'll get a thumbs up emoji on a slack thread. If i downsize it and it causes problems, i'll draw the ire of the whole team. I better leave it alone." And the truth is... by the time your company hits the end of those credits, you're probably at the point where that savings isn't gonna do much. Or maybe you are out of business.
And that is how almost every successful company that uses AWS eventually ends up paying six-figures or more annually.
It’s funny how different everyone’s experience is with this stuff. To me the diminishing returns are more around not going crazy with prototyping or running with xmax thinking all the time. I haven’t found it hard to stay under the usage limit of one $200/mo Claude and one $200/mo Codex subscription.
If my company told me yeah we’ve decided you don’t get Fable or Opus 5 because it’s too pricey, you gotta use GLM whatever, I’d be displeased.
1. Codex, Claude and others try to switch models being used at their level itself to manage the cost and outcomes
2. Now company like data bricks develops one more layer on the top of it to do the same task, of finding the base harness and applicable model
Companies like Codex and Claude are focussing/investing heavily on to ensure that people are using their harness directly or instead use APIs. Unless Databricks has some agreement in place they are violating the TOS and openly publishing an article about it. Would be interesting if openAi or Anthropic come back and claim for the API usage prices and all the savings go away.
... where in the article did they say they were using subscriptions? I'm fairly certain enterprises can't access subscription pricing in any case, they're all API costs (Anthropic doesn't support more than 150 on subscription pricing [0][1]).
Going through their harness (codex, claude) is subscription (app use) which is heavily? subsidized.
Anyone using the enterprise plan are charged the API pricing, however the article is not clear if Databricks is using enterprise plan or not which is why added the following disclaimer
Databricks is most certainly getting charged API pricing no matter what harness they are using. OpenAI and Anthropic models are so sought after right now that they set the terms even at the world's biggest companies, there is not a chance to get a special agreement for subscription pricing.
Why would it matter if foreign companies analyzed DoorDash data? Pizza deliveries to the Pentagon is all I can come up with, but that's publicly available at https://www.pizzint.watch/
I would bet my entire Polymarket balance ($0) that some military contractors have already asked AIs on the public Internet to design software for them.
This approach seems fundamentally predicated on being able to evaluate coding agents on your own code by having domain specific evals. With that knowledge, you can trust the routing logic is actually improving/maintaining perf while reducing costs.
Without the insight into agent performance, any changes like this feel like a gamble to save $$ at the cost of developer productivity
I'm actually working on building generic repo-specific benchmarks at https://stet.sh ;)
The difficulty of evaluating coding agents is indeed a really big challenge. We built evals on our own codebase and shared some information about that to allow other companies to replicate. We found our own evals correlated loosely with public generic SWE benchmarks.
In large user populations like at Databricks I think the ultimate answer will come from experimentation instead of offline evals. We are already doing this in small groups, exposing them to new candidate models and then measuring per-developer cost and perceived quality changes.
> Rapidly adopting newer, more efficient models delivers the largest cost wins of any technique.
I think the more interesting lever is the fourth they mention: token efficiency.
> By the time costly LLM inference occurs, the user's initial statement accounts for only a negligible fraction of the data fed into the AI system, meaning costs are dominated by context the user did not explicitly include.
I think there’s still lots of low hanging fruit in regards to monitoring and improving agent work. Look at your sessions. Look at how much time and context is being spent on, say, a web search returning dozens of results when one good single-pager doc would’ve been better.
100% - there's a lot to learn from traces from real-life sessions with coding tools! For example, I found it pretty eye-opening to see how wide the distribution of tasks truly is. There's also subtle things like how a poorly designed MCP API surface can cause a massive amount of token waste from the model just iterating on finding the right way to call it.
how do any of these routing approaches handle kv cache misses? Devin Fusion is the only one that explicitly addresses this, though it does so by switching models during compaction (not sure this isn't still a cache miss though)
We're going to do a followup blog detailing our routing approach soon! In short, the router takes in the task description and infers what models and harnesses are available and makes a recommendation up-front. So essentially the routing decision is made when the harness + model is kicked off and it's only changed halfway through if there's a major delta in complexity from the initial judgment. Therefore, most of the time the cache is maintained just as it would be before (this is the advantage of having a meta-harness that is actually planning all the sub-agents centrally)
Maintaining the cache is extremely, extremely important, so we're iterating fast but that's a major factor we track in the router's development. Couple things I'd look at:
1. The cache is generally reset after a compaction - this is the best time to make a switch if you want.
2. In many cases, the max duration of a cache is 1h, so if a session is being resumed after a long time, that's also a good time to re-assess the complexity.
We're iterating fast here and learning a lot! Definitely a lot to think about it in this area.
The kv cache is wiped as soon as you get your answer, cloud hosts are not going to hold the GPU memory for your entire session. You're probably referring to some agent level cache
These seem like the obvious tweaks akin to "using a cheaper hosting platform". I think the real savings come from careful context control for programmatic agents, careful tool awareness and usage to reduce thrashing, distilling workflows into deterministic processes and, moat importantly, adding friction and boundaries for non-technical users who tend to burn tokens making insane asks like "analyze all documents and give me a summary".
I've tested Omnigent superficially, attracted to its thinking around policy, governance, sandboxing, and ui. But it's still alpha at present. I forked its Polly model and got working a somewhat more complex multiagent workflow that I've also modeled in Sandcastle and Gas City but the agent broke after the next update which I would have needed to patch to maintain functionality. Subjectively I also noticed individual models seemed to be performing somewhat worse when wrapped in the platform's framework, presumably due to the extra context introduced (token use was measurably higher). Promising project that I'll revisit when it's further along and I do not doubt the outcomes Databricks claims in committedly dogfooding it.
Omnigent and OpenRouter are different in the sense that OpenRouter is where you can go to call the actual model but Omnigent is intended to be the place where you go describe the high level task to be done, and work is farmed out to various harnesses and models. Those sandboxes can themselves be using OpenRouter for capacity!
We're calling the layer coordinating harnesses "meta-harness'
Really? Because removing it from my company has saved us over 2 million a year and we were able to speed up processing. The chargeback model for databricks is predatory at best.
I think you’ve misunderstood the article. It’s about how Databricks reduced their own costs, not about how adopting Databricks will reduce anyone else’s costs.
I would be really curious to hear from devs at Databricks what the experience of development is like internally. I work at a small startup with essentially unlimited AI spend budget - the entire point is that I should be turning to it at every opportunity since our human labor is so expensive relative to tokens. So generally it's like:
- Spend most time prioritizing/discussing what to do.
- Once that's agreed, use Fable 5 High + 5.6 Sol XHigh come up with a design + plan. Agree on the high level plan. (Usually this just comes down to choosing where the change belongs on the spectrum between minimal patch <-> full redesign)
- Use Opus 5 or Sol Med to execute
- Auto-fix bugs and CI until green + thermonuclear review skill x3.
- Manual interrogation of change/nits
- Come up with QA plan and have Codex Computer Use execute on it
- Manually spot check the final result (usually a sizable diff, thousands of lines, complete feature E2E, etc)
I probably spend like $80 a day at least but I produce the output of 3 or 4 2022 engineers and probably at better quality. So it's easily worth it. Would I save money by switching to GLM 5.2 and such...perhaps? IDK. At our scale it's not worth the time spent building the eval harness to actually understand the performance tradeoff.
In our team's experience, the product of agents is generally The Homer (1). It does work, but it's vastly overengineered.
When I personally want tight code, I have to spend a considerable amount of time adjusting it manually:
- It needs to be trimmed down. In my experience, at least one agent I use struggles to produce minimalist designs, and it's very frustrating
- I need to consider whether there are solutions based on higher-level assumptions, that AIs typically miss
- I need to check whether there are off-the-shelf solutions - AIs like to reinvent the wheel
IMO, software production has become a mass-produced commodity in every sense - it's much more expensive to produce software manually, but the quality is not the same.
(1) https://simpsons.fandom.com/wiki/The_Homer
It took me a few seconds of deliberating if The Homer was a reference to baseball or "The Odyssey" and then realized there was a footnote
> software production has become a mass-produced commodity in every sense - it's much more expensive to produce software manually, but the quality is not the same.
Agreed.
This was more true a few months ago but Fable has improved the situation considerably.
Also just remember - minimalist code looks and feels great but customers do not read your code. I have caught myself many times providing "corrections" to abstractions that were already ~fine, just not perfect. The average SWE costs $200/hr. Careful you don't burn $50 worrying about code that will likely be rewritten or can be better abstracted when that's actually needed.
> The average SWE costs $200/hr
This is a pointless quibble but the hourly rate claim is not true--it's like ~$60 in the USA [0]. Maybe you meant at a specific Org but this is important context when comparing "pricing" between human and AI.
[0] https://www.salaryexpert.com/salary/job/software-developer/u...
How is this not true? Taking a Senior SWE @ ~$200K, even just the base salary cost / 2080 working hours is $100/hr. Fully loaded employer cost + accounting for non-coding time gets you to upper 100s easily.
Even for a junior making $100K, I have a hard time believe their time is worth less than $75/hr or so.
Edit: Fine, "Senior" is not "Average". But naive salary is not the true numerator.
Company time != Pay rate, if you're working somewhere that's publicly traded check out "revenue per employee" metrics sometime.
An MBA's rule of thumb is that a full time employee's hourly cost to a business is at least 1.5x to 2x times their salary depending on employer taxes, benefits, offices, travel, training, hardware, perks, etc.
$15, where we're going.
Also, you can probably think about it like this:
"Will I benefit from this code being minimalist before [date]", where [date] is whenever you think the agent will be good enough to come back and make the corrections you would make today.
Minimalist code is necessary to keep AI agents working well for longer than a month on a system IME. At a certain point, their own machinations overwhelm them and they both slow down, and make worse and worse decisions.
> The average SWE costs $200/hr.
And this is how I find out I'm woefully underpaid.
No surprise, LLM companies optimize for waste. More tokens, and more prompts means more revenue. Reminds of Google’s Prabhakar Raghavan story: deliberately making search worse [1]
[1]: https://pluralistic.net/2024/04/24/naming-names/#prabhakar-r...
IME this works until it does not. This approach works well at the beginning of a greenfield project, but at the same time because it is so easy to add features, you will likely ship something that is way too over engineered. And that complexity will not amortize over next increments and will more likely lead to the entire project being a black box only fully understood by AI. However a more careful use of AI for targeted surgical changes is far more ”productive” in the long term IMO.
Disagree. I operate this way inside a multi-million line legacy codebase.
> I work at a small startup
How does a “small startup” end up with a multi million line “legacy” codebase? Something not mathing
> How does a “small startup” end up with a multi million line “legacy” codebase?
Easy! The output of 6 months ago Opus! Which seemed so wonderful at the time.
Have you worked at many startups?
Something isn't clear about the size of your codebase here and the level of reliability your customers expect, as a reader of your comments. Clarity there will help.
My observation has been:
- Initial greenfield work by an LLM is fast and very effective with minimal or no human oversight.
- Subsequent work ends up being over engineered and very verbose. Assumptions are made that aren't suited to the problem at hand (for example I find Fable is extremely regex happy where structured data would work much better from a readability perspective.)
- Once code bloats beyond a certain point due to unguided LLM usage, complexity is high enough that only LLMs can operate on the codebase with any economical amount of time.
- Rinse repeat and your code ends up unclear about any state that's not explicitly being tested and verified in QA loops
For some of our products this has been fine, for others it's been problematic. An understanding of your size and reliability requirements will help make the conversation more productive.
> unguided LLM usage
Why aren't you guiding your LLM usage? Is that what I said - to spam it and not guide anything? Or to have a careful workflow where you agree on design and maximize your human judgement/leverage?
> any state that's not explicitly being tested and verified in QA loops
As opposed to before, when engineers perfectly reasoned about code behavior from first principals and QA was unnecessary?
> Why aren't you guiding your LLM usage? Is that what I said - to spam it and not guide anything? Or to have a careful workflow where you agree on design and maximize your human judgement/leverage?
You didn't say anything positively or negatively regarding this so I made an assumption that you were using the LLM relatively unguided (e.g. a bit of oversight, not the kind of thing that heavy code reviews used to involve pre-agents.) Feel free to add clarity on your actual usage loop.
> As opposed to before, when engineers perfectly reasoned about code behavior from first principals and QA was unnecessary?
In my experience, most engineers are quite good at reasoning about code behavior for non-QAed code paths. Obviously things fall through the cracks. But I've been in the ground floor of plenty of Big Techs in their early stages before agents and, yes, a lot of initial development had spotty test coverage and yet most of the engineers had good mental models of what was happening. It used to be a very valuable skill to wrap your head around a torrid piece of code with few or no tests but was nonetheless a core piece of your application. Conversely, agentic development can bring cognitive debt [1].
===
This isn't a fight. We aren't sparring over what's right and wrong. I'm just curious how other people use agents in their work as someone who is also now in a startup that uses LLM agents heavily and has no limitations on spend.
[1]: https://martinfowler.com/fragments/2026-02-09.html
No, but not relevant.
What is the point of working at a startup if you’re dealing with millions of lines of legacy code ? Isn’t the whole point of startups to create & innovate with a clean slate and modern tools?
No, actually. The point is to build a profitable business.
I don't know if that's the whole point, but I agree with the sentiment, why would a startup be working in legacy code and where would that code come from if this is truly the start of something.
OP might just be working at a small software company or for one that broke from a bigger one and is now "startup" like?
"startup" and "legacy codebase" are diametrically opposed concepts.
And if you're saying (based on your other comments) that a 6 month window is enough to create a legacy codebase...that indicates a serious lack of experience or understanding as to what a legacy codebase is, or why they exist.
exactly. Usually legacy code forms when people lose context and confidence in parts of the codebase due to staff turnover etc and ppl avoid touching or enhancing those parts for long periods. Six months is a short time to accrue that much tech debt, its enough time where most of the people who created that "legacy" are probably still around. As you said indicates bigger problems.
This is very close to my workflow but you forgot one important step:
- Suggest a better approach that makes the AI say, “That’s much simpler. And you’re right. My original plan was over-engineered.”
Do you have issues with performance at the moment? Right now I tend to find that it produces absolutely terrible design patterns and especially performance. I mean maybe I don't know exactly what area you're looking at but yeah for us we tend to find it's terrible wrt dB/caching/scaling and often any performance improvements it proposes end up actually shooting itself in the foot and being worse than before but it's not very good at testing in an organized way to even notice it made it worse despite repeated prompts to do so I mean if I prompt it to test performance in a handheld structured way (it is very bad at finding out what performance to test and why) before making changes I can usually figure it out but it usually takes insistence on the specifics to really ensure a good solution that will actually fix the problem
Are you using the SOTA models at very high reasoning during planning? IME that makes a LOT of a difference. I‘d also never let them just rip into the architecture, but always push back and ask for alternatives first. Once the overall plan is nailed, not that much can go wrong. Provided it’s a reasonable change set and not a 20k LOC PR.
Performance is better than ever. It's never been more practical to set up wildly complex synthetic test environments and measure perf wins. Plus the models will find every possible algorithmic/design improvement.
It actually gives me quite an uncanny feeling, bulldozing over years of human optimization work with a newer, "perfect" design. Like bringing an AK-47 back to the middle ages.
I needed to thoroughly test rerankers on my companies rather unique corpus.
Opus and I wrote a parallelized test harness and labeled groundtruth in around 2 hours.
In 2022 that would've likely been all I did for a couple sprints
Yes 100%. This morning I casually prompted Codex to drive the browser to complete extensive performance testing in-situ that would have literally been weeks of work before. Probably in reality it just wouldn't have been done, and performance guarantees would have been attempted up front via more careful design.
In this case the design was also AI generated, and there were limited wins to be found because the design was already superb.
but I produce the output of 3 or 4 2022 engineers and probably at better quality.
Possibly, but the output of a 2022 engineer is about 1/10th of the output of a 2010 engineer, so it's an extremely low bar.
also - as always with these claims there's no actual product / repo / whatever one could check.
I would love to see what these tools create but outside slop there's never: This works, is in production, here's the code.
Any day now.
Do you have tips for generating clean productive output per dollar?
in my humble experience it boils down to mastery.
Are you at least conversational in the subject matter? You're gonna have a good time just by paying attention and adjusting your workflow. If you're getting a lot of back and forth with it, its asking a lot of planning type questions, stop, step back, rethink the whole feature, and start again from the beginning with everything more fleshed out.
If you are in a brand new field, there's no way to bridge that divide. The issue is you don't know what is good or bad, or whether what you have learned is good or bad. You're in a sports car and you don't know how to drive much less what's track and what's field.
You can spend a lot of effort getting good at prompting towards writing tests and E2E tests to at least verify your app does what you expect it to, regardless of experience.
This is a great point and I agree. My own productivity varies based on what part of the codebase I'm working on. If it's "been in there before" and I know the right questions to ask, I can one-shot a good design/improvement. If I'm spending 20-30 minutes asking Fable to "draw a diagram so I can understand" - probably less so. But notably, I CAN get there in a fraction of the time it would have taken before. You can general personalized onboarding docs to ~anything.
I appreciate this non-judgmental description of what it's like to approach a topic/technology from a newcomer's perspective. Thanks!
Keep the decision-making and execution separate. Use the high IQ models to chat about the design and make them drive subagents to do the actual work. "Chat" style threads are actually quite cheap. Where it gets expensive is having Fable 5 output thousands of lines of implementation where 95% of it was already overdetermined and there were only a few important judgement calls.
I actually have no doubt that I could replace my Opus 5 Low/Medium subagent profiles with Grok 4.5/GLM 5.2/Deepseek v4 Flash and perf would probably be pretty similar.
On top of that - highly recommend adding accurate cost counters to your statusline. You can't improve what you don't measure! (Or even have any intuition about).
> essentially unlimited AI spend budget
> I probably spend like $80 a day
This doesn’t sound like “unlimited”, I spend more than this out of pocket per day and I have a strict budget.
It's a fair point, it's not truly unlimited and I do wonder how that would change my workflow. I can definitely imagine if I was inside Anthropic or OAI with unlimited "fast" tokens, you would be more tempted to hand over even more of this process. I completely understand why they talk about "graph engineering" and such, my entire workflow above could be a graph and I could try to increase my leverage even further. Realistically though I am bounded by product decision making, not code output right now.
Surprisingly pragmatic and info packed article..
Kudos to databricks, I also find it interesting that such different companies (Stripe, Ramp, Databricks) are all building the exact same internal tools.
I think building companies is going to look more generic in the future because intelligence is an API now.
There are a surprising number of articles like this along the lines of, "we started using AI tools and ended up spending millions per year".
On what planet do people start paying for things without keeping an eye on the costs and no-one notices until you have spent a crazy amount? I don't understand. You are either paying a fixed amount which you are happy about in-advance or you are PAYG in which case you would ballpark how much it costs.
Otherwise it reads a bit like a fake problem, because it didn't really happen, you just foresaw it (as you should) and added a few guide rails.
The issue is the growth rates can cause costs to drastically change quickly. If you have 1,000 employees and the average is spending $100/month you're at a $1.2M run rate. But suddenly a new model comes out that's twice as expensive, there are some changes to the harness (we found randomly Claude Code and other harnesses will make changes that drastically impact efficiency), and then maybe you have some organic user growth as well and BOOM suddenly you're at a $10M run rate within 60 days. And it's now impossible to forecast future growth.
It is true that this problem can be mostly managed by the techniques we mention here. Those are actually pretty difficult to set up at scale, so many companies (including us) we only really did this in earnest once we started to see those large cost oscillations.
The main reason we shared this here is to maybe help other companies get infrastructure in place before massive cost swings rather than after.
Something underlying a lot of this is that pricing models for enterprise coding tools have changed from seat-based to consumption-based pretty quickly, as AI usage has exploded. For months, engineers were able to use unlimited AI for no marginal cost, but that's changed quickly.
In addition, we're seeing people applying AI to more and more use cases, so token growth is very significant. Paired with consumption pricing, it's brought this problem to the forefront very quickly for lots of companies.
> we started using AI tools and ended up spending millions per year
This is how AWS made its fortune.
Not only this, but perhaps even more nefarious is that AWS gives lots of startups $100k+ in credits. This feels generous when you get it. In reality, it means that (unless you are in a compute intensive startup) you can go for months or years before you hit this, but by the time you do, you already have very solid monthly spend.
Initially, you picked the Multi-ZA RDS db.t3.2xlarge instance because you figured "eh i have credits anyway". Two years later, someone looks at this and says "hey, this is expensive and I bet we can do everything we need on a machine half the size". But then they think "if i downsize it and that works, i'll get a thumbs up emoji on a slack thread. If i downsize it and it causes problems, i'll draw the ire of the whole team. I better leave it alone." And the truth is... by the time your company hits the end of those credits, you're probably at the point where that savings isn't gonna do much. Or maybe you are out of business.
And that is how almost every successful company that uses AWS eventually ends up paying six-figures or more annually.
On this planet?
They’re not saying they regret doing it, or that it was a mistake.
They’re just saying they’ve gained experience and have leveraged the tools to an extent their usage can be optimized.
Pretty standard business or life iteration.
It’s funny how different everyone’s experience is with this stuff. To me the diminishing returns are more around not going crazy with prototyping or running with xmax thinking all the time. I haven’t found it hard to stay under the usage limit of one $200/mo Claude and one $200/mo Codex subscription.
If my company told me yeah we’ve decided you don’t get Fable or Opus 5 because it’s too pricey, you gotta use GLM whatever, I’d be displeased.
I find this funny and interesting at some levels
1. Codex, Claude and others try to switch models being used at their level itself to manage the cost and outcomes
2. Now company like data bricks develops one more layer on the top of it to do the same task, of finding the base harness and applicable model
Companies like Codex and Claude are focussing/investing heavily on to ensure that people are using their harness directly or instead use APIs. Unless Databricks has some agreement in place they are violating the TOS and openly publishing an article about it. Would be interesting if openAi or Anthropic come back and claim for the API usage prices and all the savings go away.
... where in the article did they say they were using subscriptions? I'm fairly certain enterprises can't access subscription pricing in any case, they're all API costs (Anthropic doesn't support more than 150 on subscription pricing [0][1]).
[0]: https://support.claude.com/en/articles/9797531-what-is-the-e...
[1]: https://support.claude.com/en/articles/9266767-what-is-the-t...
Going through their harness (codex, claude) is subscription (app use) which is heavily? subsidized.
Anyone using the enterprise plan are charged the API pricing, however the article is not clear if Databricks is using enterprise plan or not which is why added the following disclaimer
> Unless Databricks has some agreement in place
Databricks is most certainly getting charged API pricing no matter what harness they are using. OpenAI and Anthropic models are so sought after right now that they set the terms even at the world's biggest companies, there is not a chance to get a special agreement for subscription pricing.
They certainly have an enterprise plan?
Open ai is allowing subscription use, anthropic also paused the effort to stop subscription use.
They did? Is there a source where I can learn more? I'd love to use my Anthropic subscription with opencode.
Bans are not in effect?
Databricks will be using the API anyway, thats all you get with an enterprise agreement.
buddy, they're on enterprise plans paying per token
Careful. If you admit to using models that weren't trained by OpenAI or Anthropic then you might hauled in front of Congress: https://www.scmp.com/news/china/diplomacy/article/3362616/us...
I'd prefer congress to be asking questions (this is all they are doing so far, based on the article) before doing any legislating.
Why would it matter if foreign companies analyzed DoorDash data? Pizza deliveries to the Pentagon is all I can come up with, but that's publicly available at https://www.pizzint.watch/
I would bet my entire Polymarket balance ($0) that some military contractors have already asked AIs on the public Internet to design software for them.
This approach seems fundamentally predicated on being able to evaluate coding agents on your own code by having domain specific evals. With that knowledge, you can trust the routing logic is actually improving/maintaining perf while reducing costs.
Without the insight into agent performance, any changes like this feel like a gamble to save $$ at the cost of developer productivity
I'm actually working on building generic repo-specific benchmarks at https://stet.sh ;)
The difficulty of evaluating coding agents is indeed a really big challenge. We built evals on our own codebase and shared some information about that to allow other companies to replicate. We found our own evals correlated loosely with public generic SWE benchmarks.
In large user populations like at Databricks I think the ultimate answer will come from experimentation instead of offline evals. We are already doing this in small groups, exposing them to new candidate models and then measuring per-developer cost and perceived quality changes.
> Rapidly adopting newer, more efficient models delivers the largest cost wins of any technique.
I think the more interesting lever is the fourth they mention: token efficiency.
> By the time costly LLM inference occurs, the user's initial statement accounts for only a negligible fraction of the data fed into the AI system, meaning costs are dominated by context the user did not explicitly include.
I think there’s still lots of low hanging fruit in regards to monitoring and improving agent work. Look at your sessions. Look at how much time and context is being spent on, say, a web search returning dozens of results when one good single-pager doc would’ve been better.
100% - there's a lot to learn from traces from real-life sessions with coding tools! For example, I found it pretty eye-opening to see how wide the distribution of tasks truly is. There's also subtle things like how a poorly designed MCP API surface can cause a massive amount of token waste from the model just iterating on finding the right way to call it.
how do any of these routing approaches handle kv cache misses? Devin Fusion is the only one that explicitly addresses this, though it does so by switching models during compaction (not sure this isn't still a cache miss though)
We're going to do a followup blog detailing our routing approach soon! In short, the router takes in the task description and infers what models and harnesses are available and makes a recommendation up-front. So essentially the routing decision is made when the harness + model is kicked off and it's only changed halfway through if there's a major delta in complexity from the initial judgment. Therefore, most of the time the cache is maintained just as it would be before (this is the advantage of having a meta-harness that is actually planning all the sub-agents centrally)
Maintaining the cache is extremely, extremely important, so we're iterating fast but that's a major factor we track in the router's development. Couple things I'd look at:
1. The cache is generally reset after a compaction - this is the best time to make a switch if you want.
2. In many cases, the max duration of a cache is 1h, so if a session is being resumed after a long time, that's also a good time to re-assess the complexity.
We're iterating fast here and learning a lot! Definitely a lot to think about it in this area.
The kv cache is wiped as soon as you get your answer, cloud hosts are not going to hold the GPU memory for your entire session. You're probably referring to some agent level cache
These seem like the obvious tweaks akin to "using a cheaper hosting platform". I think the real savings come from careful context control for programmatic agents, careful tool awareness and usage to reduce thrashing, distilling workflows into deterministic processes and, moat importantly, adding friction and boundaries for non-technical users who tend to burn tokens making insane asks like "analyze all documents and give me a summary".
I authored this - happy to answer any questions.
First time hearing of Omnigent. Anyone have experience using it?
I've tested Omnigent superficially, attracted to its thinking around policy, governance, sandboxing, and ui. But it's still alpha at present. I forked its Polly model and got working a somewhat more complex multiagent workflow that I've also modeled in Sandcastle and Gas City but the agent broke after the next update which I would have needed to patch to maintain functionality. Subjectively I also noticed individual models seemed to be performing somewhat worse when wrapped in the platform's framework, presumably due to the extra context introduced (token use was measurably higher). Promising project that I'll revisit when it's further along and I do not doubt the outcomes Databricks claims in committedly dogfooding it.
First the mofos force you to use AI then they become stingy about it.
An AI-edited post by the way.
is this opensource or have to buy from Databricks?
It seems Databricks open-sourced it a while ago:
https://www.databricks.com/blog/introducing-omnigent-meta-ha...
https://github.com/omnigent-ai/omnigent
Omniagent looks quite similar to OpenRouter (https://openrouter.ai/)
Omnigent and OpenRouter are different in the sense that OpenRouter is where you can go to call the actual model but Omnigent is intended to be the place where you go describe the high level task to be done, and work is farmed out to various harnesses and models. Those sandboxes can themselves be using OpenRouter for capacity!
We're calling the layer coordinating harnesses "meta-harness'
Probably coulda got every dev a local model for how much they spent; what a brialliant set of economists
Quit cold turkey and you can drive down AI coding spend 100%.
Really? Because removing it from my company has saved us over 2 million a year and we were able to speed up processing. The chargeback model for databricks is predatory at best.
What did you move to and what type of workload, if I may ask?
I think you’ve misunderstood the article. It’s about how Databricks reduced their own costs, not about how adopting Databricks will reduce anyone else’s costs.
Yawn. Databricks and their half baked overly expensive platform.
Too bad their AI query generation is next to useless.