Antigravity Teamwork for long-running agent tasks

Logged at IST: 2026-09-01 09:01 IST

What it is: Prashanth Subrahmanyam's hands-on walkthrough of Google Antigravity's /teamwork-preview, using it to run a long multi-agent coding task.

Gist: The post is useful because it shows the product shape, not just the launch claim. Teamwork first rewrites the user's rough requirement into a spec, asks clarifying questions, defines acceptance criteria, and presents an approval artifact. After approval, it delegates to a team-lead/orchestrator setup that breaks the work into milestones and spins up implementation, challenger, reviewer, forensic auditor, and final victory-auditor subagents.

The concrete test project was Anotato: a small app for pasting screenshots, adding visual annotations, and copying an annotated image plus textual notes back into a coding harness. Prashanth says the build ran for a little over two hours, needed no further human intervention after plan approval, and produced a working deployed app and source repo. His visible workflow notes match the official Antigravity docs: Teamwork is designed for long-running projects, structured handoffs, milestone artifacts, isolated workspaces, and independent verification gates.

The broader signal is that multi-agent coding is becoming a productized harness pattern. The interesting parts are not just “more agents,” but spec refinement before execution, adaptive team composition, liveness monitoring, explicit challenger/auditor roles, and a final artifact that summarizes tests and run instructions.

Newsletter angle: Long-running agent work is moving toward managed teams: approval-gated specs, milestone plans, subagent handoffs, adversarial verification, and final “prove it works” audits.