Introduction
A research-grade agent harness for bounded coding work, verified outcomes, and explicit human authority.
Private internal project. Features described here are evidence-limited and are not offered as a public execution service.
The problem
An AI-generated patch can look convincing while failing independent tests. Repeating the same model prompt can consume more time and compute without producing a different answer. Collective Agent explores how to make each coding attempt auditable, measurable, and subject to explicit approvals.
The workflow
A task begins with a frozen source revision and defined scope. A worker proposes a change; an isolated executor runs permitted checks; independent acceptance determines whether the candidate passes; a signed receipt records the result. An operator decides whether a patch should be applied.
Current scope
The project is an internal Rust-based agent fork in active experimentation. Narrow local execution and supervised child jobs have been validated in specific test fixtures. This is not a public software service, an unattended deployment system, or a universal security certification.
How to read these docs
Start with Getting started to understand the task lifecycle, then Architecture and Permissions for execution boundaries. Verification distinguishes reported experiments from demonstrated production capabilities. The Roadmap lists unverified work.