Four Additions

Our resource catalog now lists 234 entries: three tools for recording, checking, or gating agent actions, one benchmark, and no removals.

  • HOL Guard: a local tool for supported AI coding agents that checks selected tool actions against policy, asks for approval or blocks them, and keeps decision receipts for review. Coverage varies by agent integration. Code.
  • Mycelium Trails: specifications and offline verifiers for agent-action references, SHA-256 digests of four metadata fields, which its server anchors on Arbitrum One. A third party holding those fields and the transaction reference can check when a digest was committed; the digest does not cover action payloads. Code.
  • TRACES.md: a plain-text, append-only trace format and CLI that records each operation and its result, with a validator that rejects a step marked ok that reports a non-zero exit code. Append-only is a convention, not cryptographic tamper evidence. Code.
  • EC2-Bench: 128 executable workspace tasks, 100 with privacy annotations, that measure task performance, cloud cost, and how much sensitive workspace information reaches a cloud collaborator during edge-cloud agent runs. Paper, Code, Data.

Pull requests to the companion list proposed all four: #56 (giskard09), #57 (passionate11), #58 (kantorcodes), and #59 (MaxonT), merged November 16. Each submitter is the project's author, maintainer, or a disclosed affiliate; the HOL Guard submitter disclosed a leadership role at HOL, which maintains it. Maintainer edits narrowed the Mycelium Trails entry to what its code anchors and added Rust to HOL Guard's language tags. Inclusion does not certify a tool, benchmark, or security claim. This date marks catalog publication, not project releases. Updates appear in our RSS / Atom feed.