Missions: Multi-Agent Systems That Ship for Days — Luke Alvoeiro, Factory

The limiting factor in modern software engineering is not AI intelligence but human attention and supervision bandwidth (1:00). Even the best engineers can only drive a few tasks forward per day because each requires review and oversight, while models are smart enough to handle 50 tasks simultaneously if given the right structure.

The talk proposes a taxonomy of five frontier multi-agent communication patterns (2:15) — delegation (parent spawns child agents), creator-verifier (separate build and check agents), direct communication (agents DM each other), negotiation (agents coordinate over shared resources), and broadcast (one-to-many information sharing). Missions combines four of these into a single workflow.

Missions uses orchestrator, workers, and validators (4:30). The orchestrator handles planning and scoping, workers implement features with clean context and commit via Git, and validators verify behavior through both traditional checks and end-to-end QA testing. This separation ensures no single agent accumulates degraded attention.

Validation contracts are written during planning before any code exists (6:15), defining correctness independently of implementation. This prevents the common failure mode where tests written after implementation merely confirm decisions rather than catching bugs, and enables the system to run for days without drifting.