Independent judgment
No self approval.
Each reviewer writes only its own authoritative review file. The approver reads those files and fails closed on missing or malformed verdicts.
Spec first. Tests first. Independently reviewed.
Agents build.
Evidence ships.
ANVIL / Agentic Network for Verified Implementation Lifecycle
A software factory where no agent approves its own work.
Turn a work order into specs, tests, implementation and a review trail you can inspect.
27 stations8 pipelinesCatalog verified at build
2026-11-07
2026-11-07
The problem
A convincing demo
isn't a review.
When the agent that writes the code also tests and approves it, confidence becomes a closed loop.
And when a demo leaves no spec, test evidence or run record, nobody can audit what happened. ANVIL separates the work from its judgment and keeps the artifacts in your repository.
How it works
A factory line.
Not a longer prompt.
The spec comes before implementation. Tests come before the build.
Each station has a contract and an assigned artifact. Pipelines compose those stations into a line that fits the work: software, content, a new project or an existing codebase.
sdlc-default / work flows through the line6 reviewers / parallel, independent
The checks gate runs after the builder, before model review. The reviewers check spec, security, tests, architecture, design and UX in parallel. The approver aggregates their files; the human gate stops risky work before deployment. The observer follows the release.
What makes it different
Trust has a paper trail.
Not just an answer. The evidence behind it.
Verified memory
Learnings earn their place.
Proposals stay quarantined until deterministic code checks their provenance against the run record. Only verified learnings enter station context.
Deterministic checks
Checks before review spend.
The gate runs the configured project suite and checks protected paths before model reviewers run. Failed checks return the work to the builder.
Rendered design review
Review what people see.
Interface review uses rendered captures when available, with findings tied to the page itself. Source alone cannot prove a layout works.
After deployment
Check the live result.
Configured post deploy health verification checks live routes and platform signals. An unknown result is recorded as unknown, never assumed healthy.
Cost visibility
A report, including cost.
Per run reports total measured model costs and show measurement coverage. Missing usage or prices stay unmeasured, never invented.
Proof, not promises
Follow the work.
Products built on the line. Factory work recorded in the open.
Sentia
A full-stack web app with database persistence, from spec through deployment verification.
Platform25
A playable game built through the content pipeline, with media production and brand review.
The factory's own work
Inspect work orders, station packets, independent review files and run reports in this repository.
Release v0.1.0
A versioned starting point. Read the release and inspect the code behind it.
Quickstart
Start with something
you can inspect.
Clone Taskline, the worked example. Run its tests, then explore the specs, reviews and run report alongside the code. Requires Git and Python.
git clone https://github.com/arananet/anvil-example-taskline.git
cd anvil-example-taskline
python3 -m unittest discover -s tests -vThese are the README quickstart commands. To set up your own factory line, continue with the documentation index.
What ANVIL is not
Autonomy with accountability.
Not a single agent.
Not a prompt library.
Not unsupervised autonomy.
The human gate is a feature. ANVIL prepares the work and its evidence. A human decides what gets published or deployed to production.