The /teamwork-preview slash command - working not working?

Google Antigravity
1

Multi-Agent Teamwork (Ultra Plan Only)

Antigravity 2.0 introduces advanced multi-agent orchestration for extremely complex tasks.

Ultra Plan Exclusive: The /teamwork-preview slash command is currently in preview and is exclusive to users on the Ultra plan. or is this is this in Australian dollars an thus the $345 a month plan? I may have mis read or not understood what plan i am on.

image
image1819×1330 110 KB

This is from my usage page when logged into https://gemini.google.com/usage?hl=en-AU

Using /teamwork-preview prompts the main agent to launch a collaborative multi-agent framework. This framework features built-in error recovery, automatic retries, and coordination logic, allowing you to simply define the high-level goal while the platform manages the overhead of a cooperative agent team.

i am on this plan but i can seem to get this command to work.

its looks awesome. so what gives?

cheers,

Desmack.

2

This is interesting, I was asking this myself yesterday.

The blog post about /teamwork-preview skill was created before the Ultra Plan split (from no options (x20) to two options (x5 and x20).

So yeah, we probably need confirmation from the Antigravity Team to see if this is intentional or was overlooked.

3

It seems to have been enable now.. its appeared in the slash commands

image
image1436×1458 182 KB

i will let you know how it goes.. #excited.

Cheers,

Desmack.

4

Howdy team,

Just ran your beta teamwork dash preview control command and holy shit — way too much uncontrolled exhaust. Dozens of Markdown files dumped straight into my agent’s working directory. It’s a mess.

Bare minimum fix needed:
Add a config option (or flag) so I can point all that output/exhaust to a dedicated folder instead of polluting the main agent spot. That alone would make it usable.

Second thing (strongly recommend):
Switch the generated docs to .toon format instead of .md.

Why? Here’s the actual maths (using real TOON benchmarks):

TOON (Token-Oriented Object Notation) is specifically built for LLM/agent contexts like this.

  • Official benchmarks show ~40% fewer tokens vs JSON on mixed structures, with even bigger wins on tabular/agent data (up to 60% in some cases).

  • Markdown is worse than clean JSON for token efficiency — all the #, -, **, code fences, and extra whitespace easily add 30-80% more tokens compared to a compact structured format like TOON, depending on the content.

If you’re generating 30-50+ of these files per run (like in that commit), you’re burning tens of thousands of extra tokens per cycle on pure formatting bloat. That’s real latency + cost for the orchestrator and every downstream agent.

Links for you:

Seriously, switching to .toon would cut token rot hard. Markdown is fine for humans reading on GitHub, but it’s garbage for LLM harnesses.

Appreciate the tool otherwise — just trying to help make the beta not drive people insane.

Cheers,

image
image1084×1893 424 KB

5

Final results trying the new /teamwork-preview command on a complex refactor of a heterogeneous compute pipeline (LLM inference across CPU + GPU, involving Triton, Cython, and custom TritonPath kernels).

What happened:

  • It generated ~50 Markdown files and destroyed my token count with massive uncontrolled output.

  • It spent most of its time spinning, running overly zealous test suites, and second-guessing itself instead of actually doing the work.

  • It failed to make meaningful progress on the actual refactor I requested.

For comparison, I fed the exact same instructions into the normal anti-gravity pipeline (without the teamwork preview) and it dispatched the agents and completed the work successfully.

I was really hoping the preview mode would aggressively rip apart and coordinate the complex task, but right now it feels like it’s adding a lot of overhead and noise without delivering results.

Quick suggestions for the beta:

  • Add a config option to direct all the “exhaust” (docs, briefings, progress files, etc.) into a dedicated folder instead of polluting the main agent directory.

  • Strongly recommend switching generated files from Markdown to .toon format. Markdown is extremely token-inefficient for LLM contexts.

It’s clearly still early beta, — just wanted to share real usage feedback from a heavy, non-trivial workload.

Cheers,

Desmack

6

Nice, so it was just overlooked.

Related to the formating: at that point wouldn’t it be better to store everything as tokens in a blob?