Build Relay into agent actions
Review proposed actions before execution
Relay returns a clear decision that the next agent, product, or runtime can use
First review path
Guided first review page to Workspace sandbox to production Passport
Copy the preview command first, create a live decision only when ready, then open Workspace for proof and sandbox MCP
Open first review
Copy the command
Run preview
Check the decision shape
Open Workspace
Inspect decision trace
Register Passport
Create identity
Review proof
The current proof set for Relay before action
Run local review, risk-check, review-path, and CommerceOps examples before live decisions
Relay reviews proposed actions before execution. It is not a separate product, and Relay never executes downstream work
The agent proposes the action, Relay reviews it, and your runtime keeps execution
Agent action
Proposed action first
Relay review
Checks permission, policy, and risk
Clear decision
Proceed, revise, review, or stop
Runtime executes
Your system keeps control
Run the GitHub example
Review locally first, optional MCP after
Core review, MCP path, Passport lookup, and blocked examples
Clone repo
Run local example
Run review
Decision returned
Open Workspace
Inspect decision and trace
Replay trace
Check evidence
Request / Decision / Route
Send the proposed action, get the decision, route the next step
const actionCard = { version: "0.1", agent: { id: "agent_ref" }, proposedAction: { summary: proposedAction }, affectedObject: "object_ref", context: { evidenceRefs: ["evidence_ref"] }};GitHub examples
One repo, four review paths
Public path stays direct review first; MCP remains optional protected access
Relay / Registry / Protocol
Three roles, one review path
Proposed actions enter Relay, Registry supplies participant identity, and Protocol keeps the shared record consistent
Relay
Reviews actions first
Registry
Identifies who is acting
Protocol
Keeps the shared decision record
What changes
A governed decision layer with a visible mechanism
Relay reviews the proposed action and returns a decision the next agent or system can use before execution
Agent remains yours
Your runtime keeps the agent, tools, workflow, and execution path
Decision record
Relay records who acted, what changed, evidence, policy, and risk
Routing becomes clear
Your system uses the decision to proceed, revise, review, or stop
Start with one review
Run the example, open Workspace, then create Agent Passport
Run the review example, inspect the decision in Workspace, activate sandbox MCP if needed, then create the production Agent Passport