Damascus — Self-Improving Agent Kernel on AMD

Self-improvement claims in AI are usually vibes. Damascus makes them falsifiable. A builder agent (Gemma 3 27B, self-hosted with vLLM on an AMD Instinct MI300X via ROCm) attempts a frozen benchmark of 15 small coding tasks, each graded by its own pytest suite. An improver agent reads the failures and may edit exactly one file — AGENTS.md, the project rules injected into the builder's prompt. It can never touch the tasks, the tests, or the code, and this is enforced in code with a path check, not just in the prompt. The loop then re-runs the full benchmark against the exact same tests. The key mechanism is eval-gating: if a rules update drops the pass rate, the kernel rejects it and reverts automatically — like a package manager refusing a broken dependency. In our runs, the pass rate climbed from 60% to 91% across 8 runs (14/15 tasks passing), with 6 of 8 proposed rule revisions accepted and 2 rejected and reverted with no manual intervention. Every number is reproducible from the runs/ artifacts: per-task transcripts, pass/fail JSON, and every AGENTS.md snapshot. The agent never grades itself — pytest is the only judge. No LLM-as-judge anywhere in the pipeline. The diff history between AGENTS.md versions is a human-readable record of what the agent actually learned, making Damascus the missing regression test for prompt engineering. A dashboard renders it all like an OS: process monitor, kernel version, and a ledger of accepted and rejected updates. Damascus steel gets stronger every time it's forged. So does this agent.