The Multi-Agent Architecture That Actually Ships — Luke Alvoeiro, Factory (Agents column)

I saw something interesting. video Luke Alvoeiro from Factory (theirs Droid The first place in terminal bench v1) A system where multiple AI agents conduct long engineering tasks: plan, write code, check, repair and do not lose context for hours or days. The idea is that if before the question was “can a model write code?”, now the question is, can the system autonomously bring the task to a working state and prove that the result is correct? The Factory guys call it a mission. (mission)The user describes the goal and scope, and then the system breaks the work into parts, launches agents, validates the result and returns the problems to the patch cycle.

1Multi-agent - This is not to run 10 agent Parallelism alone does not make the system smarter. Separation of roles is important: Orchestrator specifies requirements, builds a plan and validation contract. Workers implement specific parts with limited context. Validators independently verify the results. The value is not in the number of agents, but in the reliable cycle: plan → execute → validate → fix → repeat.

2Validation contract is written before the code Verification should appear before implementation, because if the agent first wrote the code and then wrote the tests himself, then these tests easily become confirmation of the decision already made. As a result, it is first formulated what should be true for the user and the system, and only then the code is written. This is similar to TDD, but at the level of the whole task.

3Validator must be independent The auditing agent should not drag the context of the author of the implementation. Otherwise, he will inherit the same assumptions and blind spots. That is, it is better to check the coding agent with another model, another session or a separate context - by acceptance criteria, not by the author's explanation. (The author is the original coding agent.)

4Write code better consistently If Just run agents in parallel, then problems arise – they change the same files at the same time, conflicts, duplicates and incompatible interfaces appear. As a result, there is a basic approach that can work well: write operations to build sequentially, and research, reading documentation, codebase search and validation can be paralleled.

5The state must live outside the context window A long task cannot be held in the memory of a single agent. External artifacts are needed: validation contract, feature list, research notes, guidelines, knowledge base. Hence, a new criterion for the quality of the repository: how much it is AI-ready. README, AGENTS.md, conventions, tests, scripts, CI and clear boundaries are becoming infrastructure for autonomous development.

As a result, a good AI workflow from the author of the report looks something like this.

1. Validation contract first.
2. Then a limited scope.
3. Structured handoff: what's done, what's not done, what commands have been launched.
4. Independent verification.
5. And then merge.

And developer value shifts from manual code writing to designing boundaries, checks, and execution contour.

#AI #Agents #Architecture #SystemDesign #Engineering #Software #Management