Sheep isn’t just an idea — it’s already running. This page outlines what we’ve built, what’s in progress, and what’s next.
Built the foundation that manages goal decomposition and interdependent tasks dynamically. Agents can already coordinate on small-scale software operations such as code edits, tests, and documentation.
Implemented a semantic indexing system that scans codebases, creates embeddings and metadata, and retrieves relevant files for each task using a local vector store (Chroma / FAISS).
Connected Sheep to GitHub for real repositories: it now generates branches, applies AI-generated patches, and opens pull requests ready for human review.
Added an approval interface where human reviewers can inspect, correct, and approve AI-generated work. All feedback is logged and used to continuously train the agents.
The execution graph can now restructure itself based on results, errors, or new priorities discovered during runtime. This allows Sheep to act like a living project manager, optimizing work as it happens.
Developing specialized “communication agents” that reach out to human teammates on Slack or email when context is missing, ensuring continuous collaboration without bottlenecks.
Experimenting with a closed ecosystem of specialized agents (code, QA, documentation, DevOps) that share context and reasoning. Early tests show promising cooperation and reduction in repetitive queries.
Gathering structured logs and review data from Sheep to train the reasoning layer that will power the STEP Protocol.
Preparing pilot programs with two software houses to validate Sheep’s efficiency in real production environments.
Building the pipeline that transforms human corrections into actionable model fine-tuning data.