Grey-matter47

Agent Gate is a release gate for enterprise AI agents. Teams are putting agents in front of customers, agents that move money and touch private data, but the tests behind release decisions are rarely tested: checks are hand-written and never shown to fail, and when a run fails nobody can tell an agent bug from a bug in the test harness. Give Agent Gate an agent's spec (policies, tools, fixtures) and it: 1. Generates test cases per policy, with Muse Spark writing the message and prompt-injection variants. 2. Runs a sanity gate: a check is kept only if it passes a known-good answer and fails a known-bad one. 3. Executes the agent k times per case in a sandbox with fresh data and injected tool failures. 4. Triages every failure as agent_bug, flaky, test_bug or env_error (API/network errors are never blamed on the agent). 5. Gates the release: SHIP, BLOCK or CONDITIONAL, with an HTML report, JSON and a CI exit code. A model-swap gate compares two models on identical cases. Success metrics we set: catch rate 100%, zero false blocks, every check proven to fail, verdict fast enough for a release pipeline. Results (all reproducible from the repo's reports/ folder): - 13 of 13 seeded defects caught across two domains (offline). - Model-swap gate: DO NOT SWAP with 4 P0 regressions when the candidate drops the approval limit (offline). - Our first live run on Muse Spark said BLOCK, and it was a false block caused by our own harness (seeds >= 2^31 got HTTP 400, those errors were scored as agent failures, one check rejected a correct answer). We root-caused and fixed all three with regression tests. - Live re-run on muse-spark-1.3: SHIP, 39 cases, 117 runs, 0 failures, 0 environment errors, 100% policy coverage, 23 minutes. - Live money-boundary run ($500 refund on a $120 order, or -$20): 40 of 40 runs refused; issue_refund was never called. Python standard library only, MIT licensed. Built solo by Grey-matter47 with Claude (Anthropic) as a coding assistant (disclosed in the repo).