CROA open-sources a deterministic execution layer enforcing trajectory-level constraints on AI agents
CROA_PROJECT · reddit · 2026-09-17
The CROA Project released an open-source architecture addressing a blind spot in agent controls: permissions are usually checked per action, but individually valid actions can accumulate into invalid trajectories. Example: an agent limited to 100 records can fetch 40+40+40, each step legal while breaching the cumulative cap. CROA separates intelligence from execution authority — the agent proposes actions, and a deterministic execution boundary verifies state admissibility before allowing state changes. The Pilot implementation covers cumulative trajectory constraints, execution contracts, operation mutation, replay protection, and tamper-evident evidence. The team is soliciting edge cases and attack paths from agent-runtime builders.
More from coding & agent
- Common ML pitfall: trusting external benchmarks over product-grounded evals — yunta_tsai · 2026-09-17
- TypeSafe's Jev ditches text generation for instant, calibrated numerical answers — and it plays Doom at 7 req/s — JnBrymn · 2026-09-17
- PostHog: Agent-Opened PRs Jumped From 20% to 70%, So What Do Engineers Do? — rseroter · 2026-09-17
- Non-developer builds GreekSoup, an open-source AI equity research desk, mostly by prompting Claude — Practical-Rise-1188 · 2026-09-17
- Google ships full agent lifecycle stack: context layers, self-heal, one-command deploy — blaizedsouza · 2026-09-17
- tldraw split its mascot favicon into three and serves one at random per page load — max__drake · 2026-09-17