Features

Everything agents need to ship

Agent Swarm is built around autonomous work: many agents, isolated environments, durable memory, and proof that the code actually runs.

Parallel agents

A whole team, working at once

The lead agent fans work out to a swarm. A planner explores while a coder builds and a tester verifies — each in its own thread, none blocking the others.

  • Many threads per project, each an independent agent
  • Sub-agents drill in for focused tasks
  • Stop or steer any agent mid-run
swarm · 4 agents running
planner
done
coder
72%
tester
48%
docs
30%
Structured memory

Context that never overflows

Beyond rules and markdown, a materialized knowledge base captures architecture, conventions, and decisions as structured records — distilled and refreshed by background agents.

  • Knowledge base of architecture & decisions
  • Automatic distillation as threads grow
  • The right context loaded per request
knowledge base
System architecturestructure
API conventionsconvention
Why we dropped Redis cachedecision
Auth & session modelstructure
Video proof

Proof it works, not just that it compiles

Every implementation task produces a Playwright recording of the working feature, with acceptance criteria encoded as executable assertions that run on every change.

  • Playwright video for each implementation
  • Acceptance criteria as runnable tests
  • Regressions caught automatically
proof · checkout.spec.ts.mp4
0:18
7 acceptance assertions passed
And more

The rest of the toolkit

Isolated sandboxes

Every agent gets its own cloud sandbox, branched from main. Compute-heavy builds and test suites never bottleneck your laptop.

Agent triggers

Configurable automations fire on tool calls, events, or schedules. Docs update when code changes; tests run when specs move.

Native git

Project is a repo, each thread a branch. User messages are checkpoints, agent output auto-commits, and Commit Thread merges or opens a PR.

Acceptance-driven planning

Before code is written, the system surfaces edge cases — refresh, resize, empty and stale data — and encodes them as executable tests.

View in your IDE

Fetch any agent's branch locally and review it in Cursor or VS Code's native diff — accept or reject per file. No extension required.

Ship your next feature with a swarm

Spin up your first project in under a minute. Bring a git repo or start from an empty one.