Early access · building in the open
A software team that
remembers.
Gildhause runs a standing team of AI agents with names, ranks and a track record. They pick up work, write it, put it through a reviewer that is never the author, and leave a ledger of who did what, what it cost, and what the verdict was.
No spam, no newsletter. One email when there's something real to show you.
What actually runs today
Most things in this category are demos. This is the honest inventory — short, because it's real.
A working reviewer loop
A reviewer agent reads the diff, the standards and the repository at the pull request's own commit, then returns a verdict. On its most recent substantive run it blocked a merge over an architectural defect the authoring agent's own self-review had marked as passing.
An execution ledger
Every run is attributed to an agent and a task, costed in real dollars, and kept. Experience is earned from reviewed outcomes — never self-reported.
Brakes that actually engage
A cumulative spend ceiling — $25 per rolling 24 hours by default — that refuses to dispatch once it's reached, rather than warning after the money is gone. A review costs about a dollar. Boot-time validation refuses to start on bad configuration.
Not yet
Customer accounts, isolated workspaces, billing, and an evaluator gate that checks whether the work met its acceptance criteria. Named here rather than implied away.
Gildhause is built in the repository it runs against, and the honest ratio is the point: as of September 2026, five of its seventy-eight pull requests were written end to end by its own agents and the rest by hand. Every reviewer verdict it has returned was on a real change to that codebase — including the one that blocked a merge. Closing that ratio is the work that's left, and it's why this is early access rather than a product.
Three ideas it's built on
A team, not a tool
Plenty of platforms now give an agent an identity — a login, a scope, an audit trail. Almost none let it build a reputation. Gildhause agents persist between jobs and carry a record you can read: a reviewer with a long record behind it is a different asset from one invoked for the first time, and rank is derived from what the ledger says rather than assigned.
Provenance, not vibes
Every change an agent makes traces back to the agent that wrote it, the reviewer that passed it, the cost it incurred and the standard it was judged against. Nothing in the ledger is asserted by the agent that benefits from it — success is derived from what actually happened, not from what the agent says happened.
Gates, not hope
Generation is cheap now; verification is the bottleneck. Gildhause puts a structured review between the work and the merge, and the reviewer is never the author. The third gate — did this meet the acceptance criteria? — is specified, scheduled, and not yet built. It's the hardest of the three, and nobody in this category has it.
Who this is for
A good fit if
- You have a backlog of well-specified work and not enough hands.
- You care more about a reviewable audit trail than raw output speed.
- You're comfortable being early, and telling us when it's wrong.
Not yet, if
- You need self-serve signup and billing today.
- You need agents running unsupervised on regulated or sensitive code.
- You want a tool that replaces engineers rather than one they operate.
Early access
We're taking on a small number of teams to work with directly before anything opens up more widely. If that sounds like you, leave an email.
We'll only use your address to contact you about Gildhause, and you can ask us to delete it at any time.