All projects
Agent runtimeIn productionSince 2025

Relay

Durable execution for AI agents that take real actions.The runtime behind our agent deployments. Relay runs multi-step agents reliably: tool calls, retries, human approvals and full traces, even when a run takes minutes or days.

relay · run r_8f21cExample trace
  1. run startedworkflow support.resolve_damaged_item
  2. ✓ orders.lookup182 ms
  3. ✓ shipments.track240 ms
  4. ✓ policy.check replacement_eligiblepassed
  5. ⏸ approval.request refund ¥12,800 > ¥10,000 limit#support-leads
  6. ✓ approval grantedafter 3m 12s
  7. ✓ payments.refundidempotency key r_8f21c:5
  8. run complete7 steps · 4.1 s active · ¥2.7
run completion rate
99.7%
run completion rate
duplicated side effects on retry
0
duplicated side effects on retry
of actions policy-checked before execution
100%
of actions policy-checked before execution
registered production tools
40+
registered production tools

Overview

Why we built Relay.

Most agent demos break in production. A tool times out, a model returns something unexpected, a step needs a manager’s approval and the process dies. Relay treats every agent run as a durable workflow. Each step is checkpointed, so runs can pause, resume and recover without repeating side effects such as refunds or emails.

Tools are registered with typed schemas, permissions and rate limits. Policies are enforced outside the model, so an agent cannot act beyond its permissions, regardless of what the model generates. Every run produces a complete trace for debugging, audits and evaluation.

Capabilities

Core capabilities.

01

Durable workflows

Checkpointed steps with idempotent tool execution, automatic retries and resume-from-failure.

02

Human-in-the-loop

Approval gates that pause a run, notify the right person in Slack or email, and continue once they decide.

03

Typed tool registry

Schema-validated tools with per-agent permissions, rate limits and spending caps.

04

Policy enforcement

Business rules checked outside the model before every action. Violations are blocked and logged.

05

Tracing & replay

Step-by-step traces of prompts, tool I/O, tokens and latency. Any run can be replayed against a new version.

06

Model-agnostic

Route between model providers per step, with fallbacks when one is degraded.

Architecture

System architecture.

A simplified view of the principal layers and components.

Built with

TypeScriptNode.jsTemporalPostgreSQLRedisOpenTelemetryKubernetes
  1. Interfaces

    REST & webhooksSlack & TeamsLINEHelpdesk apps
  2. Orchestration

    Workflow enginePlannerApproval gatesScheduler
  3. Execution

    Tool registryPolicy engineModel routerSecrets vault
  4. Observability

    TracesCost meteringReplayAlerts

Contact

Have an AI product in mind?

Share what you would like to build, and we will help turn the idea into a production-ready product.

No technical specification required.

What happens next

  1. 1

    Share your idea

    Describe the problem, the users and the outcome you need.

  2. 2

    Discovery call

    We review your requirements together and clarify scope and constraints.

  3. 3

    Proposal

    You receive a clear scope, technical approach and delivery plan.