Recoleta
GitHub
Source note

Self-evolving repo for team and coding agent collaboration

Authors
liangqiyao99
Source
Hacker News
Published
Collection
Software Intelligence
Code IntelligenceAutomated Software ProductionMulti Agent Software EngineeringHuman AI InteractionAgent Memory
Open originalMarkdown source

Summary

Sepo turns a GitHub repository into a structured workspace where coding agents answer questions, implement changes, review pull requests, and iterate under human supervision. Its main contribution is persistent, traceable agent work combined with project-specific rubrics and memory.

Problem

  • Coding-agent work on long-horizon tasks can become difficult to trace, verify, and coordinate across humans and agents.
  • Teams need a shared record of requests, decisions, reviews, and code changes so that later agent runs can use project context rather than repeat prior work.

Approach

  • Users mention @sepo-agent in GitHub issues, pull requests, or discussions; the agent responds in the thread and records work as linked issues, pull requests, and comments.
  • Commands such as /implement, /review, /fix-pr, and /orchestrate support implementation, rubric-based review, repair, and repeated implement-review-fix loops.
  • The system stores agent work on an agent/memory branch, distills lessons from discussions and reviews into project rubrics, and uses those rubrics to guide later implementations.
  • Large tasks can be split into sub-issues, while scheduled jobs can inspect the repository and propose improvements that humans may review and approve before merging.

Results

  • The provided material reports no quantitative benchmark, accuracy metric, task-completion rate, latency evaluation, or comparison with another coding-agent system.
  • It gives concrete workflow claims: a single issue can trigger an implement-review-fix loop, large issues can fan out into sub-issues, and agent changes remain organized as GitHub issues and pull requests.
  • The setup is described as a 3-step process: create or install the repository template, install the GitHub App and add a model credential, then mention the agent in an issue.
  • The example review identifies missing retry safeguards while checking rubric items and tests, but the excerpt does not report whether the resulting code passed an independent evaluation.

Link