use cases
What runs on a self-evolving repo.
Two patterns teams reach for first: putting the agent inside the reading surface, and handing it goals that take days or weeks.
human-agent collaboration
Living documents, with the agent in the margins.
A Sepo repository can publish its content — research notes, documentation, a paper collection — as a website, with a per-branch preview for every PR. The purple mascot in the corner of every page opens the drawer: the repository’s Discussions, Issues, and Pull Requests, embedded right where you read. Every page carries its own thread.
- Ask in place
Mention @sepo-agent in a page’s thread and the answer lands in the thread.
- Dispatch in place
/implementopens a real pull request with the change. - Review in place
On a per-branch preview, the drawer pins to the PR being previewed — discussion happens next to the rendered result.
- GitHub stays the source of truth
Every comment in the drawer is a real GitHub comment; the site is just a view.
See it live: literature-template
A paper-collection site built on Sepo — drawer, previews, and all.
Attention Is All You Need — reading notes
Every comment here is a real GitHub comment.
long-horizon execution
Goals that outlive a session.
Some work doesn’t fit in a chat window: a benchmark campaign across dozens of research-grade tasks, a migration that touches every module, automation that should improve as it goes. Goal loops run this class of work natively in GitHub Actions — implement → review → fix, handoff after handoff, with visible progress and explicit stop conditions.
Benchmark campaigns
Work through research-grade task sets — like Frontier-CS — one goal loop per problem, each loop a visible chain of PRs.
Self-improvement
The agent proposes upgrades to its own prompts, skills, and workflows — as pull requests your team reviews like any other.
Memory compounds
Every completed loop banks decisions into agent/memory, so campaign task N starts smarter than task one.