Descriptions:
Luke Alvoeiro, who leads core agent infrastructure at Factory and previously created the open-source coding agent Goose (now donated to the AI Agentic Foundation), presents the Missions system — Factory’s production approach to running multi-agent software development workflows autonomously for days or weeks. The talk opens with a practical taxonomy of five multi-agent communication patterns: delegation, creator-verifier, direct communication, negotiation, and broadcast. Missions combines four of these into a three-role architecture where an orchestrator defines a validation contract upfront (specifying what “done” means before any code is written), workers execute features with clean isolated context, and two classes of validators — a scrutiny validator reviewing implementation and an end-to-end validator interacting with the live application — check work adversarially.
A counterintuitive core design decision is serial feature execution rather than full parallelization. Factory found that running multiple agents simultaneously on the same codebase caused conflicts, duplicated work, and inconsistent architectural decisions whose coordination overhead erased the throughput gains. Serial execution with targeted internal parallelization on read-only operations (code search, API research, code review) turned out to produce dramatically lower error rates — and in long-running tasks, correctness compounds.
Factory’s longest mission ran for 16 continuous days — longer than a full sprint — enabled by structured handoff documents that force each agent to record what was completed, what was left undone, which commands were run, and whether it followed orchestrator-defined procedures. This written state, rather than agent memory, is what allows the system to self-heal at milestone boundaries and maintain coherence across an ecosystem of agents that never share a single context window.
📺 Source: AI Engineer · Published May 06, 2026
🏷️ Format: Deep Dive
