What it is

Factory Missions is a planning-first orchestration framework for long-running autonomous coding and adjacent knowledge work. The core move is to turn one vague multi-hour request into an explicit plan of milestones and features, then execute that plan with separate worker sessions, validation workers, and an orchestrator that can recover, redirect, and generate follow-up work.

Core mechanics

• collaborative planning before execution
• milestone-based execution structure
• fresh worker contexts per feature
• validation workers at milestone boundaries
• targeted rather than blanket parallelism
• model/tool specialization by role
• human-as-project-manager intervention loop

Operating thesis

The framework assumes that reliability comes from structure, not raw autonomy. Long-lived single sessions drift. Broad parallelism creates coordination drag. The better path is controlled decomposition, narrow worker scope, explicit checkpoints, and recovery mechanisms when workers stall or validation fails.

What seems strongest

• Treating planning as the real leverage point.
• Making validation part of the execution graph instead of a final pass.
• Using fresh contexts to reduce drift and repeated confusion.
• Admitting that operator intervention is part of the system design, not a failure mode to hide.

Main weaknesses / limits

• Economic profile may be hard to justify outside high-value work.
• Coordination overhead can become its own tax if milestones and roles are over-specified.
• Success depends heavily on plan quality and validation quality.
• Public evidence so far is still closer to breakout + docs than broad independent benchmarking.

Relevance to this wiki

Factory Missions is one of the clearer real-world examples of the wiki's broader argument that useful agent systems behave more like managed execution organizations than like a single chatty agent. It is especially relevant to pages on planner/executor separation, artifact-first handoffs, and harness-first development.