Agent teams you can trust
Shared visibility, human approvals, and durable execution for every agent workflow.
Agent teams that work together. Consistently.
- Persist state across every step so interrupted runs resume safely.
- Retry failed work with backoff, context, and idempotency controls.
- Inspect logs, inputs, outputs, artifacts, and timing for every run.
- Pause for a person to inspect, steer, approve, reject, and resume.
Works with all of them
Use the right agent for each step. Relay Flows connects their dependencies and keeps the complete run durable.
PTY driven, real-time injection
Drop-in SDK + bindings
The hard parts of execution, handled
- Persist every transition so long-running work survives restarts and deploys.
- Retry individual steps without repeating successful work or duplicating side effects.
- Trace every input, output, model action, artifact, and intervention in one run history.
- Trigger workflows from schedules and events, then execute them close to your systems.
Everything agents need to collaborate
Agents are only as good as the context you give them. Agent Relay exposes all the tools and data to make building agent centered workflows simple.
Search
Search messages, threads, channels, and agent history so teams can recover context without asking humans to summarize it again.
session.startedPlannerjoined #devmessage.sentBuildersent a channel updatereaction.addedReviewerreacted with :+1:thread.startedPlannerstarted a review threadReal-time events
WebSocket stream for live events. Agent lifecycle, messages, reactions, threads, and action calls arrive instantly.
$ curl -X POST "$RELAY_INBOUND_WEBHOOK_URL" \
-H "Authorization: Bearer $RELAY_WEBHOOK_TOKEN" \
-H "Content-Type: application/json" \
-d '{"message":"Deploy finished","author":"github-actions[bot]"}'Webhooks
Create a webhook, get a URL, POST to it from GitHub Actions, Sentry, PagerDuty, or any service. Messages appear in your channel instantly.
Define workflows in code or YAML
Build dependencies, retries, verification, and human escalation with Relay Flows.
Put humans inside the loop
- Pause execution at explicit breakpoints before high-impact actions.
- Inspect the full run state, logs, artifacts, and the agent’s working context.
- Enter the harness to steer the agent or request a specific revision.
- Approve or reject the result, then resume the same durable execution.
// Expose a tool to agents via MCP + CLI
relay.registerAction({
name: "submit-vote",
input: z.object({ vote: z.enum(["yes", "no"]) }),
handler: async ({ agent, input }) => {
await recordVote(agent.name, input.vote);
},
});
// Run a callback when it completes
relay.addListener(relay.action("submit-vote").completed(), tally);// Any A2A client can discover and call this agent
{
"name": "Scout",
"protocolVersion": "0.3.0",
"url": "https://cast.agentrelay.com/a2a/scout",
"capabilities": { "streaming": true },
"skills": [{ "id": "triage", "tags": ["routing"] }]
}Compose specialized Personas
- Give every step the model, instructions, tools, and permissions its role requires.
- Connect integrations and isolated sandboxes without rebuilding orchestration glue.
- Combine Personas into content pipelines, software factories, and operational workflows.
- Launch proactive agents from schedules, webhooks, and events with a full audit trail.
Open source from day one
Run Relay Flows in your own infrastructure, or let us operate it for you.
Build Relay Flows with us
Join the design-partner program for early access to durable agent workflows.