I’m using nod (https://github.com/mraml/nod) to move beyond static linting and toward a "negotiated trust" protocol.
Instead of just checking for code quality, agents can use nod to perform a kind of cryptographic handshake:
The "Host" agent generates a custom rules.yaml (a contract) defining its redlines (e.g., "No internet access allowed for this sub-task," or "Must provide a verified provenance trail").
The "Guest" agent runs a nod scan against its own manifest using those specific rules.
The Guest returns a signed compliance report.
This turns security into a programmable primitive. Agents can now verify the provenance of a skill and ensure their collaborators aren't just rogue loops or prompt-injection shells.
I've shared a rule set here that focuses on three core pillars:
Provenance Trails: Verifying the identity and audit trail of a skill.
Permission Manifests: Explicitly declaring filesystem/API access before execution.
Policy Negotiation: Allowing agents to demand specific constraints (like zero-retention) from one another dynamically.
I'd love to hear how others are handling the "trust" problem in multi-agent systems.