Related: I've been finding a lot of success (defined as more autonomy with less quality dropoff) with team-based workflows.
A lot of evidence for this.
- Adaptive Auto-Harness: Sustained Self-Improvement for Agentic System Deployment on Open-Ended Task Streams (https://arxiv.org/abs/2606.01770) talks about multi-agent, role-based adaptive harnesses collaborating on non-homogeneous workstreams
My experience using oh-my-claudecode as orchestrator and using the teams options was a mix. For one side the work completed was very good and it took less time. The main issue I had was my hard drive was getting full because I was using worktrees and the weekly tokens were spent in about 2 days (plan Max 20x)because all the context was being duplicated in all the agents as part of the teams.
Not sure how Antigravity works, but I hope it doesn't have the context issue.
As noted by others, we already announced parts of /teamwork in Antigravity here: https://antigravity.google/blog/teamwork-when-ai-becomes-a-r... with more visualizations and more. Thi post was meant to report some advances in math/science/coding/engineering to show-case examples of tasks you can do with it.
We are constantly improving/evolving /teamwork and pushing those changes for external users. Will also post more about improvements and details for building teams of agents dynamically as Antigravity blog posts soon. Stay tuned!
Can you please let the product team know that there's space between Pro and Ultra pricing?
I can't afford/allow for ultra pricing for personal projects, but Pro runs through in about two days of casual usage for me. I've paid for extra credits, but my god are those not worth the price.
I would love to pay for a sub that's between Pro and Ultra.
Second pro sub is the usual middle ground, same as with other providers. You can switch to second account using an alias (agy2='unshare ... exec agy' to remap $HOME/.gemini to e.g. .gemini2).
Although using another provider for the secon sub is probably a better idea since having a different model doing review would uncover more issues than asking the same model that generated the code..
Here is a short list of example results reported on the post; again we will have dedicated Antigravity posts for some of these.
* Math Open Problems & Formal Verification: Reporting 10+ open math and CS theory problems solved including 50M-line Lean 4 proof of Courtade–Kumar conjecture plus progress on several open problems in quantum query complexity, Tarjan's Deque Conjecture.
Also autonomous Lean formalizations of the Weil (1.8M lines), Birkhoff, and 3D Bochner–Riesz conjectures via Teamwork-LEAP, plus updates on Deep Review for STOC’27.
* Chemistry & Discovery: 95% top-1 exact match on NMR molecular structure elucidation (nearly 2× single-agent). Also empowering Nuclear Fusion and Quantum algorithm results as well autonomously generating a CausalAnalyst Benchmark. In addition, we show how /teamwork can start from Alphazero paper, implement a competitive go game, and autonomously improving it by playing with KataGo to get to the medium Elo level.
* Several Systems, TPU Kernel, & Infra:
* Clustering 1-Trillion-edge graphs on FarArray and RDMA (up to 180× faster),
* 800K-line Zircon & 170K-line libgav1 safe Rust rewrites (20–30% faster than C++/ASM),
* 4.09× JAXBench TPU kernel speedups with DuoDiscover,
* Google internal examples: surpassing 1 PiB fleet memory savings, and 24% fewer Borg TPU preemptions for Borg Scheduling.
FWIW, since Argon 4 was launched I used Antigravity (without the /boost or the /teamwork flags) for "playground research" and had some discoveries in cryptography[1] and speeding up Python, removing the GIL, improving the JIT, and optimizing it for certain libraries (e.g. NetworkX) [2]. I don't know how stable is the Python one but it passes the included tests. I have spent some GC budget on it though.
I think if you just keep focusing on ridiculously generous usage limits compared to Anthropic in order to harvest the use cases of the 99 percent, like making shitty websites, but which benefits people like me, you will find success.
Both take a long time. It used to be that antigravity was much faster than codex, but with recent releases it is substantially slower.
Inspecting the agents communication in teamwork, there seems to be an overwhelming amount of ceremony: BRIEF.md, specialized roles, multiple rounds of checking.
Is all that necessary? Codex also uses agent “swarms” but unless they’ve hidden how the sausage is made, it doesn’t seem so heavyweight.
It is not clear to me how or if /teamwork is any different than Claude dynamic workflows.
Antigravity as a harness feels very behind. There's not even any auto mode. And they announced /goal long after I ever felt the need to use it in CC or Codex.
There is (--mode auto), but you can't access it without restarting the harness. Even then, you'll get plenty of mother-may-I showstoppers. I had to resort to using the --dangerously-skip-permissions option to get anything done. The good news is that Carbon is a pretty awesome model once you do get it working.
Seems likely that the people who maintain Antigravity don't use Antigravity to do it. It has too many rough edges that would be trivial to fix if someone actually noticed them.
I'm not really sure. I've tried the EGM plugin too, but the issue that “the auditor verified it” doesn't solve the auditor's own reliability problem and just circulates the responsibility for verification within one layer—honestly, I'm curious how they solve that. What I want to know is this kind of thing.
In practice, when I use the Antigravity Argon multi-agent model, the problem of auditors making false completion reports still exists. I wish the Antigravity team would write something like a whitepaper on how they actually use the Gemini models. It works properly in other models' harnesses, and I pay for credits and use Gemini, but I've never gotten results that good from Gemini.
The amount of breakthroughs this could deliver. Previously we should’ve established R&D departments, prepare multi-cultural teams, all to work on a single project. Today? I can see a domain expert, guiding antigravity and delivering fast.
I think the authors also used Antigravity to write this article, because it sucks at communicating whatever point it’s trying to make. I think it’s something like “We did the OpenAI 100 problems thing, but across many disciplines, and in collaboration with the experts instead of just dumping results on the world, and we also made the tooling feature available,” but I’m not sure because it has that distinct Gemini flavor of packing in so many low-level details that the actual point becomes entirely obscured.
It is a shame that Gemini prose has become too detail dense in recent releases. Its prose was one of its distinct advantages over Claude with that annoying Claude speak...
Not just overly detailed, it misses important hints that would help the reader and puts several things into one sentence in a way that the sentence is difficult to follow:
> For n-bit integer multiplication, multi-agent campaigns simplified the theoretical Turing-machine bound below n log n, while a separate effort gave a new O(n log n) algorithm on a single two-dimensional grid whose C implementation matches the GNU Multiple Precision Arithmetic Library (GMP) bit-for-bit up to a billion bits.
Is "simplify a bound" a normal wording? What does it mean to simplify a bound?
It seems like they already announced /teamwork earlier [0], but now just announced that they got some scientific results out of it. They probably intentionally want to keep this low key, but in that case it's not clear why announce it at all.
It sounds like they're saying they now offer an option, "/teamwork", that will throw more computer at a problem by choosing a team and a workflow that you the user don't have to (or get to?) know about.
Interesting that a Rust re-write pre optimisation was 30% slower than the C++ equivalent.
I wonder if they optimised the C++ it would actually have more speed benefits than the Rust alternative. Only they can't claim it's a Rust rewrite like every other company at the moment.
I’m not sure if they are that far behind given the integration forced into every single one of their consumer products but yeah their developer LLM focus has been very disappointing.
I still don’t understand why they don’t use their massive market position to give more value to Google workspace customers to have more access to their models for teasing higher value plans. The existing credits timeout before they even build a playwright scraper.
Related: I've been finding a lot of success (defined as more autonomy with less quality dropoff) with team-based workflows.
A lot of evidence for this.
- Adaptive Auto-Harness: Sustained Self-Improvement for Agentic System Deployment on Open-Ended Task Streams (https://arxiv.org/abs/2606.01770) talks about multi-agent, role-based adaptive harnesses collaborating on non-homogeneous workstreams
- team-solve (https://github.com/guregodevo/memdoor/tree/main/examples/wor...) team of small models beats Fable on LiveCodeBench
- Agensh paper (https://arxiv.org/html/2609.26781v1) also talks about "role-playing agents to study autonomous cooperation"
All show that role-based, multi-agent setups improve performance.
My experience using oh-my-claudecode as orchestrator and using the teams options was a mix. For one side the work completed was very good and it took less time. The main issue I had was my hard drive was getting full because I was using worktrees and the weekly tokens were spent in about 2 days (plan Max 20x)because all the context was being duplicated in all the agents as part of the teams. Not sure how Antigravity works, but I hope it doesn't have the context issue.
As noted by others, we already announced parts of /teamwork in Antigravity here: https://antigravity.google/blog/teamwork-when-ai-becomes-a-r... with more visualizations and more. Thi post was meant to report some advances in math/science/coding/engineering to show-case examples of tasks you can do with it.
We are constantly improving/evolving /teamwork and pushing those changes for external users. Will also post more about improvements and details for building teams of agents dynamically as Antigravity blog posts soon. Stay tuned!
Also I added a list of examples here: https://x.com/mirrokni/status/2121351383236436247?a=ma9UPZg_... and https://lnkd.in/p/esVTpR6p. I'll add a comment with examples as well.
Can you please let the product team know that there's space between Pro and Ultra pricing?
I can't afford/allow for ultra pricing for personal projects, but Pro runs through in about two days of casual usage for me. I've paid for extra credits, but my god are those not worth the price.
I would love to pay for a sub that's between Pro and Ultra.
Second pro sub is the usual middle ground, same as with other providers. You can switch to second account using an alias (agy2='unshare ... exec agy' to remap $HOME/.gemini to e.g. .gemini2).
Although using another provider for the secon sub is probably a better idea since having a different model doing review would uncover more issues than asking the same model that generated the code..
Neither one has privacy (model training off) so it’s pretty much just a toy or for open source folks whose codebases get stolen by Google either way
Here is a short list of example results reported on the post; again we will have dedicated Antigravity posts for some of these.
* Math Open Problems & Formal Verification: Reporting 10+ open math and CS theory problems solved including 50M-line Lean 4 proof of Courtade–Kumar conjecture plus progress on several open problems in quantum query complexity, Tarjan's Deque Conjecture.
* Chemistry & Discovery: 95% top-1 exact match on NMR molecular structure elucidation (nearly 2× single-agent). Also empowering Nuclear Fusion and Quantum algorithm results as well autonomously generating a CausalAnalyst Benchmark. In addition, we show how /teamwork can start from Alphazero paper, implement a competitive go game, and autonomously improving it by playing with KataGo to get to the medium Elo level.* Several Systems, TPU Kernel, & Infra: * Clustering 1-Trillion-edge graphs on FarArray and RDMA (up to 180× faster), * 800K-line Zircon & 170K-line libgav1 safe Rust rewrites (20–30% faster than C++/ASM), * 4.09× JAXBench TPU kernel speedups with DuoDiscover, * Google internal examples: surpassing 1 PiB fleet memory savings, and 24% fewer Borg TPU preemptions for Borg Scheduling.
Is CausalAnalyst the causal inference benchmark it talks about? Where are the results?
FWIW, since Argon 4 was launched I used Antigravity (without the /boost or the /teamwork flags) for "playground research" and had some discoveries in cryptography[1] and speeding up Python, removing the GIL, improving the JIT, and optimizing it for certain libraries (e.g. NetworkX) [2]. I don't know how stable is the Python one but it passes the included tests. I have spent some GC budget on it though.
[1] https://arxiv.org/abs/2611.00512
[2] https://github.com/srw/cpython-experimental
I think if you just keep focusing on ridiculously generous usage limits compared to Anthropic in order to harvest the use cases of the 99 percent, like making shitty websites, but which benefits people like me, you will find success.
When to use teamwork vs boost vs swarm?
Both take a long time. It used to be that antigravity was much faster than codex, but with recent releases it is substantially slower.
Inspecting the agents communication in teamwork, there seems to be an overwhelming amount of ceremony: BRIEF.md, specialized roles, multiple rounds of checking. Is all that necessary? Codex also uses agent “swarms” but unless they’ve hidden how the sausage is made, it doesn’t seem so heavyweight.
It is not clear to me how or if /teamwork is any different than Claude dynamic workflows.
Antigravity as a harness feels very behind. There's not even any auto mode. And they announced /goal long after I ever felt the need to use it in CC or Codex.
Auto mode has been there for few versions, and recently (1-2 days ago?) became the default
There is (--mode auto), but you can't access it without restarting the harness. Even then, you'll get plenty of mother-may-I showstoppers. I had to resort to using the --dangerously-skip-permissions option to get anything done. The good news is that Carbon is a pretty awesome model once you do get it working.
Seems likely that the people who maintain Antigravity don't use Antigravity to do it. It has too many rough edges that would be trivial to fix if someone actually noticed them.
I read changelog every release, and there's always a lot of work updating auto mode in last series (1.5.x)
This doesn't tell me how to use it.
I'm not really sure. I've tried the EGM plugin too, but the issue that “the auditor verified it” doesn't solve the auditor's own reliability problem and just circulates the responsibility for verification within one layer—honestly, I'm curious how they solve that. What I want to know is this kind of thing.
In practice, when I use the Antigravity Argon multi-agent model, the problem of auditors making false completion reports still exists. I wish the Antigravity team would write something like a whitepaper on how they actually use the Gemini models. It works properly in other models' harnesses, and I pay for credits and use Gemini, but I've never gotten results that good from Gemini.
The amount of breakthroughs this could deliver. Previously we should’ve established R&D departments, prepare multi-cultural teams, all to work on a single project. Today? I can see a domain expert, guiding antigravity and delivering fast.
I think the authors also used Antigravity to write this article, because it sucks at communicating whatever point it’s trying to make. I think it’s something like “We did the OpenAI 100 problems thing, but across many disciplines, and in collaboration with the experts instead of just dumping results on the world, and we also made the tooling feature available,” but I’m not sure because it has that distinct Gemini flavor of packing in so many low-level details that the actual point becomes entirely obscured.
It is a shame that Gemini prose has become too detail dense in recent releases. Its prose was one of its distinct advantages over Claude with that annoying Claude speak...
Not just overly detailed, it misses important hints that would help the reader and puts several things into one sentence in a way that the sentence is difficult to follow:
> For n-bit integer multiplication, multi-agent campaigns simplified the theoretical Turing-machine bound below n log n, while a separate effort gave a new O(n log n) algorithm on a single two-dimensional grid whose C implementation matches the GNU Multiple Precision Arithmetic Library (GMP) bit-for-bit up to a billion bits.
Is "simplify a bound" a normal wording? What does it mean to simplify a bound?
Wow, that passage is garbage. This "bit-for-bit" stuff is the newest LLMism, by the way ...
It seems like they already announced /teamwork earlier [0], but now just announced that they got some scientific results out of it. They probably intentionally want to keep this low key, but in that case it's not clear why announce it at all.
[0]: https://antigravity.google/blog/teamwork-when-ai-becomes-a-r...
I'm pretty sure they didn't because of the following nonsensical sentence and how generally disorganized the post was:
"In under eight days, the engine gained roughly 2,600 Elo, reaching 1,885 Elo, placing it between KataGo at 16 and 32 visits per move."
It sounds like they're saying they now offer an option, "/teamwork", that will throw more computer at a problem by choosing a team and a workflow that you the user don't have to (or get to?) know about.
> I think the authors also used Antigravity to write this article, because it sucks at communicating whatever point it’s trying to make.
The pattern seems to be pervasive at Google. Too much Ai, not enough human, top-down orders
Interesting that a Rust re-write pre optimisation was 30% slower than the C++ equivalent.
I wonder if they optimised the C++ it would actually have more speed benefits than the Rust alternative. Only they can't claim it's a Rust rewrite like every other company at the moment.
I also thought of what speedup the tool could produce for the C++ original
Google is missing LLMs the way that Microsoft and Intel missed cell phones.
I’m not sure if they are that far behind given the integration forced into every single one of their consumer products but yeah their developer LLM focus has been very disappointing.
I still don’t understand why they don’t use their massive market position to give more value to Google workspace customers to have more access to their models for teasing higher value plans. The existing credits timeout before they even build a playwright scraper.
Sloppification is now called "advancing". That may even be innovative.