Multi-Agent Software Engineering System

A design for coordinating several narrowly scoped language-model agents on one engineering change, with a deterministic orchestrator that owns the plan, the state and every transition between steps.

Single-prompt coding assistants lose track of large changes. Engineering work has distinct phases — understanding, planning, implementing, verifying, reviewing — and each benefits from a different context and a different definition of done.

A deterministic orchestrator owns the task graph: a planner decomposes the goal into steps, a researcher gathers repository context, an implementer proposes changes in a sandboxed workspace, a tester runs the project's own checks, and a reviewer judges the diff against the plan. State lives in the orchestrator, not in agent memory, so any step can be replayed.

Language-model agents with narrow roles and tool permissions; the orchestrator — not a model — decides transitions, retries and termination from test results and budgets.

Agents never merge their own work: a failing check or a reviewer rejection sends the step back, and consequential actions stop at a human checkpoint. Token and time budgets are enforced per step.

Design objective: every accepted change is traceable to a plan step, passes the repository's own quality gates, and can be replayed from recorded state.