Pilot now open · limited customers
The control plane for production AI workflows

Steer long-running AI workflows
without starting over.

Relay lets your team change a running agent workflow in real time — preserving completed work, avoiding recomputation, and keeping a full audit trail. Think Temporal for agentic systems, with the live observability of Datadog.

Self-host in under 10 minutes. Your data and models stay in your environment.

Drop-in control plane · works alongside LangGraph & CrewAI · no workflow rewrite

Real-time steering Durable checkpoints Incremental recompute Full audit trail Approval controls
The problem

AI workflows don't fail because models are wrong.
They fail because the world changes mid-run.

A long-running agent workflow can take minutes or hours — approvals, tool calls, data jobs. Halfway through, a policy changes, a human steps in, or priorities shift. Today, that means one thing: cancel and start over.

💸

Restart is expensive

Restart-based systems re-run every completed step on each change — paying again for the same compute, tokens and tool calls, and waiting all over again.

🧩

State gets lost

Intermediate results, partial progress and context are discarded on cancel. There's no clean way to resume from where the work actually was.

📉

The audit trail breaks

When a run is killed and restarted, the record of what changed and why is fragmented — a problem for regulated and high-stakes operations.

Why frameworks aren't enough

Agent frameworks orchestrate. They don't let you intervene.

LangGraph, CrewAI and custom loops are great at defining and running a workflow. But they treat human input as something that happens before or after a run — not during it. There's no production-grade layer to safely change a workflow while it's executing, and prove what you changed.

Define the workflowframeworks do this
Run to completionframeworks do this
Change it while runningunsupported
Preserve completed work on changeunsupported
Prove every mutationunsupported
The platform

Steerable workflows, not restart-based ones.

Relay models your workflow as a live, mutation-aware graph. When a change arrives, it applies only to the affected steps — everything already completed stays done. The simple version: your workflow can change its mind without losing its memory.

When something changes mid-runRestart-based systemsRelay
Completed stepsre-run from scratchpreserved
Recomputationall stepsonly affected steps
Execution statediscardedcheckpointed & resumable
Mid-run interventionnot possiblefirst-class & live
Record of the changefragmentedfull before/after/delta
Unsafe changesno guardapproval-gated, fail-safe
How it works

Six steps, no rebuild required

Conceptually simple — built for CTOs, product leaders and ops owners, not just engineers.

1

Workflow starts

You run your workflow through Relay — directly, or wrapped around an existing agent loop.

2

Checkpoints are created

As each step completes, its state is durably checkpointed so progress is never at risk.

3

A change arrives mid-run

A human, a policy, or an upstream signal asks for a change while the workflow is still executing.

4

Only affected steps change

Relay applies the change to just the impacted part of the workflow — not the whole thing.

5

Completed work is preserved

Everything already finished stays finished. No re-running steps that didn't change.

6

Audit trail is updated

The change is recorded with its before, after and exact delta — exportable as one record.

Measured impact

An observability surface for steering, not guessing.

Relay ships with a live control-plane dashboard. Every figure below is computed from your own workflow runs. We deliberately show definitions here — your real numbers appear in your pilot.

Recomputation avoided
measured per run
Share of completed work that did not have to be re-run after a mid-run change.
Why it matters: directly cuts compute, token and tool-call spend on every change.
Steering success rate
measured per run
Portion of valid steering instructions that were safely applied as a graph change.
Why it matters: shows the system reliably acts on operator intent.
Mean recovery time
milliseconds
Time from a detected failure to a recovered, continuing execution.
Why it matters: bounds downtime and protects your SLAs.
Human turnaround time
measured per run
Time from a human instruction to a produced decision in the running workflow.
Why it matters: measures how fast people can actually course-correct.
Audit completeness ratio
measured per run
Share of mutations carrying a full before-state, after-state and exact delta.
Why it matters: compliance, provability and clean incident review.
Runtime stability
errors · drops · crashes
Dropped events, runtime errors and crashes observed during execution.
Why it matters: the baseline of production trust for any control plane.
Integration

A drop-in layer over the stack you already run.

Python SDK LangGraph CrewAI Copilot / Cursor / Claude workflows Custom agent loops Self-hosted
  • Install the lightweight Python SDK and point it at your Relay control plane.
  • Run native Relay workflows, or wrap an existing agent loop with a few lines.
  • Adapters for LangGraph and CrewAI map your existing task graph in.
  • Use Relay around Copilot, Cursor or Claude-driven product-build workflows when specs, guardrails or approvals change mid-run.
  • Steer, checkpoint and audit any workflow — no rewrite of your logic.

Works as a drop-in control plane layer — no need to rebuild your workflows.

from relay_sdk import RelayClient

relay = RelayClient("https://relay.yourco.internal")

# Wrap a workflow you already run — no rewrite required.
ex = relay.start(scenario_id="invoice_approval")

# Steer it mid-run. Relay applies only the affected change.
relay.steer(ex.execution_id,
  "Add a second sign-off for payments over $50k")

# Read measured impact.
k = relay.kpis(ex.execution_id)
print(k.steps_preserved, k.steering_success_rate)

# Or bring your LangGraph pipeline straight in:
from relay_sdk.adapters import LangGraphAdapter
LangGraphAdapter().submit("support pipeline",
  ["triage", "draft reply", "issue refund", "notify"])
Where teams use it

Built for workflows that change halfway through

💳

Finance operations

What changesA payment crosses a threshold and suddenly needs one more approval.
What Relay doesAdds that approval step without throwing away the work already finished.
Why restart hurtsWithout Relay, the whole invoice flow may need to be run again from the start.
🎧

Customer support escalation

What changesA refund now needs a manager check before it can go out.
What Relay doesReroutes the live case and keeps the research and draft work already done.
Why restart hurtsRestarting means agents and humans repeat the same context gathering again.
🧠

ML & data pipelines

What changesA stage has to be redone, paused or rolled back after the job already started.
What Relay doesTouches only the affected stage and keeps the finished stages intact.
Why restart hurtsTeams lose hours of completed data prep or training time.
🛠️

AI coding agents

What changesThe spec changes while Copilot, Cursor or Claude are already helping build, test or review a project.
What Relay doesAdds new checks, approvals or tasks around that work without restarting the whole build flow.
Why restart hurtsYou lose progress across code generation, tests, reviews and handoffs that were already completed.
Return on investment

Turn every interruption from a re-run into a few seconds.

Relay converts system behavior into business outcomes: lower compute spend, faster recovery, reduced operational risk, and clean compliance records. Here's how the math works on a single workflow line.

↓ Compute
recomputation avoided on every mid-run change
↓ Time
recovery measured in seconds, not full re-runs
↓ Risk
unsafe changes gated; nothing applied on a failed decision

Worked example — illustrative model

Assumptions you can replace with your own: a 20-step workflow; ~12 completed steps re-run per interrupt under a restart model; ~$0.40 compute/API per step; ~6s per step; 3 course-corrections per day; 250 working days.

Re-run cost per interrupt (12 × $0.40)$4.80
Wasted wait per interrupt (12 × 6s)72 sec
Per day (3 interrupts)$14.40 · 3.6 min
Per workflow line, per year≈ $3,600 compute + ≈ 15 hrs recovered

This is an illustrative model using placeholder rates to show how savings scale — not a benchmark and not a measured result. Plug in your real step counts and rates. In your pilot, the dashboard reports the actual measured recomputation avoided and recovery time from your own runs — never simulated. Savings scale with the number of concurrent workflow lines.

Trust & reliability

Engineered to be safe in production.

🔒 Durable checkpointing

Every step's state is written to durable storage with atomic writes, so progress is captured continuously and can't be half-saved.

♻️ Crash recovery

An interrupted or crashed run reloads its last checkpoint and continues from exactly where it stopped — no lost progress, no manual replay.

🧾 Immutable-style audit

Each mutation is appended to the trail with its before-state, after-state and exact delta — a complete, exportable record of what changed and when.

🛡️ Fail-safe by default

High-risk changes are gated behind approval. If the decision engine can't safely decide, it changes nothing rather than guessing — no unsafe mutation is ever applied.

FAQ

What buyers ask first

Does Relay replace LangGraph or CrewAI?

No. Relay is a control plane that sits alongside your existing framework or agent loop. You keep your workflow logic; Relay adds live steering, checkpointing and an audit trail on top.

Can Relay work with Copilot, Cursor or Claude-based build workflows?

Yes. Not by replacing those tools inside the editor, but by wrapping the larger workflow around them: code generation, testing, review, approvals, deploy checks and handoffs. If that workflow changes halfway through, Relay helps you update it without restarting the entire process.

Where does our data go?

Relay is self-hosted in your environment. Your data and the model you use for decisions stay with you — nothing is sent to a third party by default.

How fast can we adopt it?

Self-host in under 10 minutes, then wrap an existing workflow with a few lines of the SDK. No rebuild of your logic is required.

Are the dashboard metrics real?

Yes. Every KPI is computed from your actual workflow runs. When something hasn't been measured, Relay reports it as unknown instead of inventing a number.

What's in the pilot vs. on the roadmap?

The pilot includes the live runtime, durable checkpoint/resume, real-time steering, approval controls, measured KPIs and the full audit trail. Enterprise add-ons such as SSO/RBAC and alerting integrations are on the roadmap — we're candid about this during your evaluation.

Pilot access

Request a demo or join the waitlist

Tell us a little about your workflow. Every submission emails our founding team directly, and we follow up to schedule a hands-on walkthrough.

One form, two intents: demo request tries founder outreach immediately, waitlist stores interest for rollout updates.

We review every request personally and schedule demos manually — no auto-sequences.

Stop starting over

Give your AI workflows the ability to change their mind.