Roadmap
This page says what the library can do today and what comes next. The library is built in public, one station at a time. Each step ends with something you can run, and each has a write-up explaining the ideas behind it. Nothing below is a promise of a date.
The first delivery is complete. It is the core contract and one real station.
| You can | Read |
|---|---|
| Build a factory of your own, run a mission through it and approve the result | Getting started |
| Run the software factory on a mission, from Python or a terminal | The software factory |
| Declare typed stations that run agents, or none | Stations |
| Judge their output with plain code over recorded evidence | Gates |
| Join stations into a factory with bounded rework | Factories and rework |
| Wait for a person to approve, settle an escalation or answer an agent’s tool call, and resume from the store in another process | Approvals and escalation |
| Commit to a mission, or turn signals such as a bug report into one | Missions and signals |
| Write a way of working as a capability that a station enforces | Methodologies |
| Have work reviewed by fresh goldfish, revised in the author’s own conversation | Elephant Goldfish |
| See a run as a trace | Observability |
| Score a methodology against doing without it | Measuring a methodology |
What “complete” does not mean: the API may still change before 1.0. The getting-started path is run against a real model by hand before each release, with the live test marker. The evaluation set has been run against a real model once, on 16 November 2026.
Each of these is one stack of small pull requests and one write-up. Build, Quality and Release each have a coming-soon page.
| Step | What it adds | What you will be able to run |
|---|---|---|
| Build | A second station: a coding agent that implements an approved design in an isolated workspace. | A design in, a working change and its tests out. |
| Quality | Checks and review as gates on the build, with rework from Quality back to Build. | A factory that fails a build whose tests do not pass and sends it back. |
| Release | A publisher that opens a stack of pull requests on a code host. | An approved change delivered for a person to merge. |
| Durable runs | Adapters that run a factory on a durable engine: a local one first, a server-based one second. | A run that survives a crash and carries on where it stopped. |
| Learning | Outcome signals fed back into intake, and methodologies compared on real work. | A factory whose results become its next work. |
| A second factory | A factory in another domain, built from the same core. | Proof of which parts are truly generic. |
A way to send missions, questions and approvals through a chat tool arrives alongside Build or Quality, as another source of signals.
Not planned
Section titled “Not planned”Some things are left out on purpose.
- Autonomy by default. A person approves. The library records who, what and when. Stations that act without approval will always be a choice you make explicitly.
- A workflow engine. A factory runs in-process. Durable execution will come from adapters to engines that already exist.
- An agent framework. Pydantic AI is the agent framework. This library is the layer above it.
- Self-modifying factories. A factory may propose a change to a methodology from what it has learned. It does not apply one without a person.
How the work is done
Section titled “How the work is done”Every change is a small pull request with its own tests and its own documentation page, so it can be read as a developer would read it. The decisions page records why the foundations are shaped as they are. Contributing explains how to take part.