Observation

A coding agent can be effective inside one bounded task, but larger software work has a different problem: the project must survive across sessions, workers and changing context.

If the only source of truth is a chat transcript, important decisions and task boundaries eventually disappear into context.

Design direction

Represent work explicitly:

  • Project — the long-lived goal and shared state
  • Task — a bounded unit of execution
  • Artifact — outputs that can be inspected and reused
  • Worker capability — what an agent or external coding runtime can actually do
  • Validation — evidence that the task reached the expected state

Why it matters

Once state is explicit, a project-control agent can delegate work without pretending that every worker shares one giant context window.

It also becomes possible to resume work, inspect failures and learn from completed tasks without rebuilding the project from conversation history.

What changed

The architecture shifts from “one smart agent with many tools” toward a small software organization with explicit responsibilities and durable state.