How the lead agent works
One lead agent per thread — it assesses complexity, plans when needed, and builds with the full toolset.
Each thread has one interactive lead agent. There is no mode picker. The agent has the full toolset on every turn and decides how much ceremony the work needs.
How it chooses a path
| Complexity | What happens |
|---|---|
| Trivial | Implements directly — a typo fix, a one-line change, a quick question. |
| Moderate | Sketches a brief approach in chat, then builds. |
| Complex | Writes a living plan as a versioned document in your Plans collection. You review it on the Project Map or in Collections, then click Build. |
Plans aren't throwaway chat. They're documents the system keeps honest — re-checked as main moves, updated when the ground shifts, one click from becoming a thread that opens a PR.
What the lead can do
The lead agent can read and write files, run commands, drive the UI, spawn sub-agents, and attach or create plan documents. Heavy work — builds, test suites, Playwright runs — happens inside isolated sandboxes, not on your machine.
Streams and recovery
Each turn creates a completion stream — a single execution attempt. Streams end when the agent finishes, you stop it, it errors, or it loops. The raw message log is written to disk after each stream, so a refresh, crash, or server restart never loses your history — the conversation is reconstructed from that log, not from memory.
You can stop mid-stream and continue later; continuing starts a fresh stream that picks up where you left off.