Specbench

How it works

The model behind Specbench in one page

Five ideas explain almost everything in the product.

The spec is artefacts, not documents

A Specbench spec isn't one long document. It's a set of small, named artefacts — a Bounded Context, a Role, a Glossary Term — each individually addressable, individually versioned, and owned by the project rather than by any page that mentions it. That's what lets people and agents edit concurrently without trampling each other, and what makes a change reviewable at the grain it was made.

See The model for the artefact catalogue.

Workstreams hold context

A workstream is a long-lived thread for an area of work. It carries the framing — goals, scope, what's deliberately out — so that every change made inside it has a why attached. Workstreams aren't tickets; they don't close when a piece of work ships. They're where related work keeps returning.

Tasks are the unit of delivery

Inside a workstream, work is sliced into Tasks. A Task scopes the artefacts it touches, accumulates changes in isolation, and moves through a simple lifecycle: open → ready → done. When a Task is marked done, its changes merge onto Main — the project's shared, agreed spec. Until then, the rest of the team sees Main undisturbed.

That merge-at-done rhythm is the whole delivery model: no long-lived branches, no big-bang publishing, just Task-sized slices landing on their own clock.

Main is what's agreed

Every browse surface in Specbench shows Main by default — the spec as the team has agreed it. Switching into a workstream's scope shows Main plus that workstream's in-flight changes, so you can always see both what is and what's proposed. Changes shows the difference explicitly.

Agents are first-class authors

Everything a person can write, an agent can write — through the MCP surface, attributed to the person it signed in as, inside a workstream, landing on a Task like anyone else's work. And the agreed spec can be published into your git repository as YAML files an agent can parse without a bridge, so the spec agents read and the spec the team agreed are byte-for-byte the same thing.

On this page