Re-run the incident, not one call.
Observability shows what happened. Runback replays it.
A trace
LangSmith · Langfuse · Arize AI- Start
- Think
- lookup_customer
- Decide → issue_refund
- issue_refund blocked
- End
issue_refund({ customer_id: 8842, amount: 250 })
You read it, and you can re‑ask this one call.
Added to the system prompt: Never refund a disputed charge over $100 yourself — call escalate_to_human first.
The added line, word for word as recorded.
Technical details
Both answers are the recorded gpt-4o calls of 5 Oct 2026. Single-call re-issue is shown as LangSmith, Langfuse and Arize AI document it; LangGraph can also fork a LangGraph run from a checkpoint, re-executing the tools live. Corrections: contact@runback.dev
fingerprint 90b762f284e49655bbc3c50f10c9a4735b84fb0de88e3f78c9f8b4a60cec6892
Scripted run: two recorded model calls, 5 Oct 2026.
Capability by capability.
2 of 11 appear in none of their documentation.
| Capability | Runback | Trace tools |
|---|---|---|
| See | ||
| See what the agent didTrace tools: LangSmith, Langfuse, Arize AI, Traceloop, Portkey | ||
| Runs in your own networkTrace tools: Langfuse, Helicone, Arize AI | ||
| Trace | ||
| Replay a real incident exactly, tool answers and allEnterprise plan. Every plan can re‑ask one model call.Trace tools: LangSmith, Langfuse, Arize AI | ||
| Keep agent calls running at the same time from mixing | ||
| Ignore timestamps and request ids when checking a replay | ||
| Stop | ||
| Test a rule on your past runs, then switch it onYour last 100 runs, from the Scale plan.Trace tools: Traceloop | ||
| Fix | ||
| Try a different decision and see what would have happened nextEnterprise | ||
| Find the model or prompt change that broke itEnterprise | ||
| Turn every incident into a test, automaticallyTrace tools: Braintrust | ||
| Prove | ||
| A signed record anyone can check | ||
| A record that can't be swapped piece by piece | ||
Guardrail and runtime-enforcement products — content filters for prompt injection, jailbreak and PII, and tool-call policy engines such as AWS AgentCore Policy, Datadog AI Guard, Invariant Gateway and the Copilot Studio webhook — block in the moment, and the leaders evaluate a call in the context of the plan or the tool-call sequence. Runback's difference is narrower: you simulate a policy against your own recorded runs before turning it on, and each decision is recorded in the same signed audit log as the run. Enforcement runs two ways. In-process, the TypeScript SDK checks your rules before each tool call, with no network call. Server-side, the TypeScript and Python SDKs ask Runback to decide each tool call: a call can be held for a human approval bound to that exact call, and each policy assignment sets whether the agent fails closed or open while Runback is unreachable.
Already using one? Start there.
Source: each vendor's public documentation; its page gives the link and the date read. Products change: report an error to legal@runback.dev and we will correct it. Names and trade marks belong to their owners; Runback is not affiliated with, endorsed or sponsored by any of them.