the self-evolving repository
Let agents own your repo.
Mention @sepo-agent anywhere on GitHub and it answers, implements, reviews, and fixes your code. Every exchange lands as structured work — issues, PRs, and comments you can reference, search, and share across teams. And it grows with you: the agent memorizes project context, extracts your rubrics, and improves the next run.
Add retry with backoff #13
opened by sepo-agent
why a self-evolving repo
Sepo makes agent work structured and verifiable for long-horizon tasks.
Interact with coding agents in issues, PRs, and discussions
- Agent work becomes traceable and supports easy cross-references to past sessions via issues & PRs
- Shared by the team — humans and agents read the same record
- Code changes are guardrailed by the harness — so no surprises
The agent learns your rubrics and accumulates memory
- Agent work and runs are stored on the agent/memory branch
- Sepo distills shared lessons into rubrics from your past discussions and code reviews
- Future implementations are steered by those rubrics to reduce wasted user effort
The agent self-improves the repo with easy supervision and verification
- Scheduled jobs have agents regularly self-inspect the code and come up with improvements
- Orchestrate long-form agent tasks in PRs and sub-issues for easy human supervision and inspection
why a self-evolving repo
Sepo makes agent work structured and verifiable for long-horizon tasks.
Interact with coding agents in issues, PRs, and discussions
- Agent work becomes traceable and supports easy cross-references to past sessions via issues & PRs
- Shared by the team — humans and agents read the same record
- Code changes are guardrailed by the harness — so no surprises
why a self-evolving repo
Sepo makes agent work structured and verifiable for long-horizon tasks.
Interact with coding agents in issues, PRs, and discussions
- Agent work becomes traceable and supports easy cross-references to past sessions via issues & PRs
- Shared by the team — humans and agents read the same record
- Code changes are guardrailed by the harness — so no surprises
The agent learns your rubrics and accumulates memory
- Agent work and runs are stored on the agent/memory branch
- Sepo distills shared lessons into rubrics from your past discussions and code reviews
- Future implementations are steered by those rubrics to reduce wasted user effort
The agent self-improves the repo with easy supervision and verification
- Scheduled jobs have agents regularly self-inspect the code and come up with improvements
- Orchestrate long-form agent tasks in PRs and sub-issues for easy human supervision and inspection
Any issue or PR is a prompt — just write a comment.
Mention it with a question
Tag @sepo-agent in any issue or PR. The 👀 reaction lands in seconds, progress streams into one live comment, and the answer arrives in the thread.
👀 in seconds, progress in one live comment, then the answer.
/implement — request to pull request
Describe what you want in the issue. Sepo opens a fresh branch and fires a new PR that implements it, linked back to the request.
Add retry with backoff to API client #13
sepo-agent · fixes #12
The result is a pull request, linked back to the issue that asked.
/review and /fix-pr — finish an existing PR
On any open PR: /review scores it against your rubrics; /fix-pr picks up what's missing and pushes the fix.
Add retry with backoff to API client #13
sepo-agent · fixes #12
- clean separation of retry policy from the client
- tests cover the happy path and timeouts
- missing max-attempts guard in retry.ts
Rubric checks plus a code-quality review — verdict: minor issues.
/orchestrate — loop until it's done
A focused issue runs implement → review → fix-pr until the rubrics pass — then the loop stops and requests your review.
Harden the API client against flaky networks #14
sepo-agent · round 1 · fixes #21
Sepo is dispatching follow-up automation.
Every hop is dispatched in the open — next: review.
/orchestrate a large task — fan out sub-issues
A big issue gets broken into focused sub-issues, each launched as its own PR and orchestrated on its own loop.
sub-issues
Sepo is dispatching follow-up automation.
Sepo is dispatching follow-up automation.
Then the follow-up — the next round is dispatched.
goals & self-improvement
From user-driven implementation to agent self-improvement.
As Sepo learns from your interactions, it can start self-improving the codebase with scheduled jobs: it looks through the code and finds areas of improvement automatically, or optimizes for explicit goals set in the repo. Again, every change is tracked and organized in structured form — issues and PRs. You can review and approve each change manually, or optionally let the agent self-approve and merge.
Goal: harden api-client reliability
get started
Three steps to a self-evolving repo.
Start with the setup guiderecommended
The best way to start: the Sepo setup guide walks you through everything below in one flow.
Use the template
Create a new repository from self-evolving/repo — or install Sepo into an existing one.
Install the app, add a key
Install the Sepo GitHub App, then add a model credential —
OPENAI_API_KEYorANTHROPIC_API_KEY— as an Actions secret.Mention the agent
Open an issue and write @sepo-agent. The 👀 reaction means it’s already working.