Stacks/Coding agent

Factory Missions

Factory's Missions architecture splits large builds into fresh, narrowly scoped agent runs with a contract-first validation loop. An orchestrator plans, workers implement with TDD, and independent validators judge the result. A published Slack-clone mission ran 16.5 hours, spent more than a third of that time on validation, and closed 81 issues through repeated milestone rounds.

Use case

A system for running large software projects as many focused agent runs — orchestrator, workers, and independent validators — rather than one long session that overflows its context. The post walks through a real mission that produced a Slack clone

Category

Coding agent

Team size

not stated

Orchestrator / framework

An orchestrator investigates requirements, writes a validation contract first, decomposes work into features grouped by milestone, and steers execution. A programmatic runner spawns a fresh worker per feature (tests first, then implementation). After each milestone, fresh scrutiny validators and user-testing validators review work; the orchestrator files fix features and the milestone re-validates. Roles are separated so the orchestrator does not implement or validate directly, workers do not declare final correctness, and validators do not implement fixes. The authors say model choice is local to each role (planning, execution, or skeptical review) but do not name the models.

Memory

State is externalized into shared artifacts: a validation contract, a feature list, research notes, operational guidelines, and an evolving knowledge base. Each agent reads only what is relevant to its current job.

Evals

Validation is built into the loop: a validation contract of testable behavioral assertions, written before features are defined; workers write tests before code; independent validators exercise the system as a black box and review implementation quality. In the Slack-clone mission: 16.5 hours, 185 runs, 778.5 million tokens, 38.8k lines (52.5% tests), 89.25% statement coverage, 81 issues surfaced, 21 fix features, and every milestone converging in 2–4 validation rounds.

Observability

The post reports mission-level telemetry — runtime split, milestone waterfall, token usage, coverage, issue counts, and trajectory lengths — for the example run. A named third-party observability product is not mentioned.

Monthly cost

not stated

Models

not stated

Observability tools

not stated

Other tools

agent-browser (named as a validation tool in the contract example)Shared markdown/JSON state files (validation-contract.md, features.json, AGENTS.md)

Source

https://factory.com/news/missions-architecture

Theo Luan · Apr 9, 2026