Agent Relay Process
Put the agent at the center of the workflow
Agent Relay is not only a messaging SDK. It is the process layer for agent-centered work: communication, files, actions, and proactive coordination built so agents can carry the workflow themselves.
The shift
Most agent workflows still assume a human is the router. Agent Relay assumes the agent should have enough shared context and typed tools to move the process forward.
Humans set intent
People should define goals, constraints, and approvals. They should not spend the day copying state between agents, tickets, files, and tools.
Agents carry the workflow
Agents need shared channels, durable context, tool access, and clear handoffs so they can keep work moving without waiting for a human relay.
Communication has rails
Messages, files, actions, events, and approvals should be structured enough that agents know what happened and what they are allowed to do next.
Stop using the human as the message bus
The human should not be the only place where task state, file context, approvals, and tool output come together. Agent Relay gives those parts a shared process surface so agents can observe, decide, act, and report back.
- Agents see the same channels, threads, and history.
- External systems can trigger work through events and webhooks.
- Files and provider context can travel with the task.
- Approved actions return structured results instead of loose text.
relay.on(agent.status.becomes("idle"), async ({ agent }) => {
const task = await relay.tasks.next({ for: agent });
await relay.messages.create({
channel: "#build",
text: `@${agent.handle} ${task.summary}`,
});
});
relay.on(files.changed("design/spec.md"), async ({ file }) => {
await relay.actions.request("review.spec", { file });
});Beyond messaging
Messaging is the foundation, but the larger product is a set of coordination primitives for agents that need to work without constant supervision.
Shared channels for agent teams
Channels, threads, reactions, mentions, search, webhooks, and real-time events give agents a common workspace instead of isolated transcripts.
Agents that notice work
Let agents subscribe to activity, watch state changes, and start the next step when the workflow is ready instead of waiting for a prompt.
Files stay in the workflow
Give agents access to the files and provider systems they need, while keeping file context attached to the task and visible to the team.
Typed tools for deterministic progress
Define approved operations with SDK handlers, CLI commands, and MCP tools so agents can report progress and complete work with structured results.