Govern every agent action before it happens
AgentSafe is the gate: before a governed agent acts, it returns allow / observe / block / escalate against the agent's mandate, enforced Standards and SOPs — deterministic, signed, and anchored as evidence.
The problem
AI is scaling faster than governance
Agents are deployed across every function, but leaders can't see what they're permitted to do, can't prove what happened, and can't stop them. Governance lives in spreadsheets, not in the path of the action.
AgentSafe
A deterministic gate on every action
AgentSafe issues each agent a mandate (a VC bounded by scope, spend caps and merchants), evaluates every governed action against enforced Standards and owner SOPs with a deterministic policy engine — no LLM in the decision — and returns allow / observe / block / escalate. Rules change live, with no redeploy of the agent.
What AgentSafe does
Agent Mandates
A VC-issued mandate: scope, per-transaction & total caps, merchant allow-lists.
Standards & SOPs
Compose policy from deterministic atoms; edit live, no redeploy.
The Authorize Gate
allow / observe / block / escalate on every Ed25519-signed action; fail-closed.
Zero-dep Guard SDK
@metamynd/agentsafe-guard wraps any tool so it runs only when the gate allows.
Human-in-the-loop
An escalate verdict parks for the owner to approve or deny.
Signed Evidence
Every decision is signed and anchored on Hedera (Merkle-batched).
Trustless MCP Guard
A service re-verifies an agent's signed request against its policy bundle.
Trustless A2A Guard
A receiving agent re-verifies a caller's signed request before a delegated task runs — no handshake needed.
Payments (x402)
Pay bound to a specific authorization; two-phase authorize → capture.
How it fits together
AgentSafe runs the Agentic Governance Protocol (MAGP): a deterministic policy-core evaluated at the gate and, cooperatively, at the edge — with identity, evidence and payment bound in.
Where it applies
- Cap what an agent can spend, per transaction and in total
- Block out-of-policy actions and escalate high-risk ones to a human
- Change a rule in the dashboard and have every agent obey it live
- Let a counterparty verify an agent trustlessly before acting
Works with your stack
What good looks like
action gated
allow / observe / block / escalate
by default
See AgentSafe in action
Book a demo, or open the live product in the app to see it establish trust across your AI systems.