Open source runtime

Agent Foundry

A runtime that keeps AI coding agents on task: plan, execute, verify, with skills, agent roles and verification gates.

Problem
Coding agents drift: they over-comment, over-install, and skip verification.
Constraint
Local-first, no cloud account, no telemetry, one Python package.
Decision
Encode the workflow as skills, 3 agent roles, and 20 verification gates run from a single script.
Result
41 original skills and 3 agents, with a routing evaluation at 85.8% top-1 and 99.1% top-3 across 113 cases (82 dev, 31 holdout).

Limits

No hosted service and no guarantee that a skill helps every task. Two optional skills carry their own Apache-2.0 licence. The routing corpus is small and English-only.

Reproduce
python scripts/check.py --gate route-eval --json, then compare evals/routing/baseline.json.