What vanguarstew is really building: a measurable AI maintainer

vanguarstew is becoming the first measurable, public, self-improving AI software maintainer. Three things, each individually rare, that together nobody else has:

Why not "the best agent framework"

Because that's the wrong category. LangGraph, AutoGen, CrewAI, the OpenAI Agents SDK, the Microsoft Agent Framework — these are excellent orchestration tools. Comparing vanguarstew to them is like comparing an inference engine to a deep-learning framework: a category error. vanguarstew isn't plumbing for building agents. It's a specialist at one hard, specific task — maintainer judgment: knowing what a codebase wants next, what to merge, what to triage, when to cut a release.

Nobody has a popular benchmark for that, because measuring maintainer judgment is genuinely new. Issue-resolution benchmarks like SWE-bench measure whether a model can write a patch that passes hidden tests — a real and valuable thing, but a different one. Predicting the direction a senior maintainer will take a project is a separate skill, and it's the one vanguarstew is built to measure and optimize.

The proof we're building toward

The goal isn't a number you have to trust us on. It's a verifiable demonstration:

That's the whole point of building on real history: the receipt is public and independently checkable. You don't have to believe our benchmark; you can verify the prediction yourself.

The flywheel

This is why the contribution-scoring mechanism matters so much. It isn't bureaucracy — it's the engine:

• Contributors improve the agent's real predictive accuracy — how often it names the modules, kinds of work, and releases that maintainers actually produced next, on repos it has never seen.
• The benchmark scores each change, anti-gamed and on held-out targets, so only genuine improvement counts.
• The leaderboard shows verifiable maintainer-foresight climbing over time.
• That climbing record, plus the frozen-repo demonstration, is the proof — which draws more contributors, who push the accuracy higher.

Every merged improvement makes the agent measurably better at anticipating real maintainers, and that improvement is public and checkable. That's the loop.

What's next

Two milestones make this concrete:

• M7 — a legible foresight metric. Turn the internal composite score into a single, objective accuracy figure an outsider instantly understands and can verify against real history — did it predict the actual modules, kinds, and releases — and make it the leaderboard's headline.
• M8 — the verifiable demonstration. A clean, fair, reproducible frozen-repo prediction on a well-known repository — freeze commit, model, and cutoff stated up front — plus the public record of the metric climbing over a real track of merged, genuinely-improving pull requests.