I am working on a project called pi-env where i try to establish a safe, repeatable, and manageable agentic coding environment for the pi coding agent.
The presented Coordination Repository Pattern is a side-effect of me trying to make agentic coding experience manageable in a different, more agent-agnostic way.
I felt we could put agentic coding better 'under control' by making all project knowledge information related to its implementation available in a separate git repository as a structured cross-referenced set of requirement, decision, issue, TODO,... items readable by a human and AI-agent alongside project's implementation repository.
It turns out that the concept works quite well (for me).
Actually the pi-env itself is fully backed by its own coordination repository.
Also what felt like a big overhead work around requirement/issue management in a human world, gets done really easy in the agent world.
This initial implementation of Coordination Repository Pattern allows for the following workflow:
- In a new session with an agent-architect we define what change in the project we want, which ends with 'architect' preparing requirement items and a set of implementation issues in a coordination repo.
- Afterwards, when i confirm coordination changes, a kind of agent-trio (agent-developer, agent-reviewer and agent-tester) is spawned in a serial loop through which opened implementation issues from coordination are handled and code implemented.
Note that each agent session of a serial loop starts empty (read cheap) each time and that handled issue content is the entry point for the agent (beside its role specification) to do the work.
Issue item also works as a communication medium between agent-roles, which remains as a track of issue handling and you can see explicitly how an item has been pushed back and forth between a developer and a reviewer for example.
For anyone to be able to re-implement this kind of agentic engineering framework the Coordination Repository Pattern has been abstracted out of the pi-env project into its own github repo for you to try out, evaluate and judge.
I'd especially like to hear about what could be wrong or missing with this approach as well as any other comment and suggestion you might have.
The coordination repository pattern github repo includes pattern's position paper, pattern specification, comparison to related practices and alternatives and a link to pi-env reference implementation.
Note! Presented documents for Coordination Repository Pattern are AI-assisted.
Hi HN,
I am working on a project called pi-env where i try to establish a safe, repeatable, and manageable agentic coding environment for the pi coding agent.
The presented Coordination Repository Pattern is a side-effect of me trying to make agentic coding experience manageable in a different, more agent-agnostic way. I felt we could put agentic coding better 'under control' by making all project knowledge information related to its implementation available in a separate git repository as a structured cross-referenced set of requirement, decision, issue, TODO,... items readable by a human and AI-agent alongside project's implementation repository.
It turns out that the concept works quite well (for me). Actually the pi-env itself is fully backed by its own coordination repository. Also what felt like a big overhead work around requirement/issue management in a human world, gets done really easy in the agent world. This initial implementation of Coordination Repository Pattern allows for the following workflow:
- In a new session with an agent-architect we define what change in the project we want, which ends with 'architect' preparing requirement items and a set of implementation issues in a coordination repo.
- Afterwards, when i confirm coordination changes, a kind of agent-trio (agent-developer, agent-reviewer and agent-tester) is spawned in a serial loop through which opened implementation issues from coordination are handled and code implemented.
Note that each agent session of a serial loop starts empty (read cheap) each time and that handled issue content is the entry point for the agent (beside its role specification) to do the work. Issue item also works as a communication medium between agent-roles, which remains as a track of issue handling and you can see explicitly how an item has been pushed back and forth between a developer and a reviewer for example.
For anyone to be able to re-implement this kind of agentic engineering framework the Coordination Repository Pattern has been abstracted out of the pi-env project into its own github repo for you to try out, evaluate and judge. I'd especially like to hear about what could be wrong or missing with this approach as well as any other comment and suggestion you might have.
The coordination repository pattern github repo includes pattern's position paper, pattern specification, comparison to related practices and alternatives and a link to pi-env reference implementation.
Note! Presented documents for Coordination Repository Pattern are AI-assisted.
Thanks.