Axiom

Axiom — autonomous software-engineering swarm that builds its own skills Axiom is a swarm of nine specialised AI roles that solves Terminal-Bench 2.0 tasks by synthesising its own skill library through a learning-and-iteration loop. The Architect identifies useful abstractions per-task; the swarm synthesises them as loadable skills mid-execution; the Iterator refines weak skills across attempts; the Librarian indexes them via vector retrieval. 158 skills, all generated during the hackathon — but the durable contribution is the loop itself: point Axiom at any new benchmark or hard software task and it builds the skill library for that domain. Result. Terminal-Bench 2.0 pass rate: 94.4% (awaiting official validation). The current public top is OpenAI's Codex on GPT-5.5 at 82.0%; Anthropic's own Claude Code sits at 58.0%. Why Opus 4.7. The model stays on track over long-running agentic work. I wrote a detailed Axiom spec, fed it to Rookery, and the swarm built itself — headless agents in parallel git worktrees, dispatched and merged automatically. And the intelligence: 4.7 cracked Terminal-Bench tasks I'd previously been unable to solve myself. The infrastructure. Building Axiom required three new primitives — all OSS, all written this week: - Rookery — parallel-session runtime; the same tool that built Axiom now runs its experiments - Raven — SQLite-backed message bus for live cross-session coordination - Roost — OAuth load balancer across Max plans https://github.com/0xDarkMatter/rookery https://github.com/0xDarkMatter/raven https://github.com/0xDarkMatter/roost