The maintainer agent that maintains itself.
vanguarstew reviews, scores, merges, and closes pull requests on its own repository — autonomously, in public, with a human holding the keys. And it gets measurably better every week.
The maintainer loop
Not "did it write good code?" — "would it have steered the repo the way a strong maintainer did?"
Given a repository frozen at a moment in time, the agent decides what a strong maintainer would do next — in four steps.
Infer the philosophy
Read the history, README, and recent activity to work out where the project is going. The hardest, most important part.
Read the situation
Open issues, open PRs, recent commits, releases — the maintainer's surface as of that moment, nothing from the future.
Plan & decide
Propose the next actions and the concrete calls — merge, request-changes, triage, review, release — each with its reasoning.
Implement
Produce an actual patch when that's the right move — but writing code is only one of the calls a maintainer makes.
Proof, not a pitch
It found a flaw in its own judgment — and fixed the system underneath it.
Run live against a real, fast-moving public repo, vanguarstew caught itself being confidently wrong — no crash, no bug report. Then it did the thing that separates a maintainer from a bot.
It kept predicting a release was imminent when the maintainers cut none. Contributors sent patches to tune the model. But reviewing them, vanguarstew reached a different conclusion: the code was never the problem — the benchmark's own yardstick was.
It was scoring over "the next 5 commits" — ~24 minutes of work on a fast repo, ~46 days on a slow one. So instead of tuning the model, it redesigned the measurement itself: each repo judged over a time window matched to its own release cadence. Fix the system, not the symptom.
Every number is checkable against real git history — that's the point of building on it.
It maintains itself
The maintainer agent, running the maintainer loop on its own repository.
The same agent the benchmark scores now co-maintains vanguarstew live — openly, as a transparent bot, with the human owner holding admin and a kill-switch.
Dogfooding, in the open — the strongest proof the concept works.
How it's scored
A benchmark built from real GitHub history.
Freeze a repo at time T, let the agent plan the next moves, then judge them against what the maintainers actually did — hardened against leakage and judge bias.
Rebuild the past
Reconstruct the repo, issues, PRs and releases exactly as they were knowable at T — nothing leaks from the future.
Agent decides
The agent infers the philosophy and plans the next maintainer actions with its reasoning.
Pairwise + anchor
A dual-order pairwise judge weighs trajectory and reasoning, blended with an objective anchor of hard ground truth.
One score
Judge and anchor fold into a single tunable composite — averaged over many freeze-points and repos.
The benchmark scores the agent across many repos at once, on a per-repo time horizon and human-era history. Generalization is the whole ballgame: an agent that maintains one repo is a script; one that maintains any repo is a maintainer.
A different category
Not an agent framework. Not an issue-resolver.
Measuring maintainer judgment is genuinely new — a separate skill from wiring up agents or writing a patch that passes hidden tests.
Live evidence
Every perf:* label, measured — never claimed.
One bar per pull request, on a fixed anchor release — every real
agent/ score the maintainer bot has ever published, including the
ones that got closed. No line chart, no averages to squint at.
Where it's headed
From one maintainer to an orchestra of agents.
The judgment core proven here becomes the manager — a maintainer that doesn't just decide, but directs a team of specialists, reviews their work, and sends back rework.
vanguarstew
Sets priorities, delegates, reviews results, approves or sends rework — the maintainer's judgment, running the room.
Today: the maintainer judgment core, measured and proven. Next: it runs the team.
Powered by Gittensor
Built by a network that gets paid to improve real software.
vanguarstew runs on Gittensor — a Bittensor subnet (SN74) that rewards a network of contributors for making real, merged improvements to open-source repositories. Every merged PR makes the agent — and the benchmark — better.