The black box for AI agents. Replay the exact decision, prove why it happened, and stop it happening twice.
runback.cassette/v1 · signed · verify it yourself — no account →
Observability assumes determinism: same code, same trace, twice. Agents don't — every decision turns on context assembled at runtime: retrieved documents, tool outputs, the exact messages[]. A log shows the outcome; only a re-executable run shows the reasoning. Keep Langfuse, LangSmith and Datadog — Runback is the layer above them.
Swipe to see all columns →
Verified September 2026 — reviewed quarterly, and every ✓ above runs green in public CI ↗.
Full comparison — every tool, every capability →Forty-three seconds, no narration: open the failed run, replay it against another model to see where the decision forks, write the rule that blocks it, verify the sealed chain.
A decision record for every run: hash-chained, sealed the moment it happens, verifiable without us and producible on demand to any auditor. The portable format is the Runback cassette.
Below is a real audit export from the incident walked through on How it works.Verify it here — no account. The check runs on our server and keeps nothing; each hash checks against the one before it, and the signature establishes who produced it. Want it off our infrastructure entirely? Download the same record and run npx @runback/verify sample-cassette.json.
{
"$schema": "runback.audit/v2",
"manifest": {
"run_id": "run_a3f1b90c",
"generated_at": "2026-06-14T02:47:13.950Z",
"event_count": 6,
"algorithm": "sha256-chain",
"content_digest": "2e91e3f5fc451d2c9841d9e3b6c68d78a9b2fca00fdff98989512704180a9189",
"replay": {
"cassette_digest": "4376f4529252da46ea7da2f5a0afb59f5cb3511b13459ef8f67cf848f7a02cf8",
"entry_count": 4,
"algorithm": "oracle-chain/sha256",
"note": "The run's oracle stream — every model response and tool output, hash-chained. Re-execute the agent against `events` with @runback/replay and the cassette digest must reproduce. This is proof the recording IS the deterministic input stream of the run — an audit you can re-run, not just read."
},
"signed": true,
"signature": {
"alg": "Ed25519",
"value": "d6ab935b3ec9d96883ccb689d5589a50c985fa7109f1e321724a2623d417d7bce7f01a8707a5e11848d7a97c759370deee37297d8f6f3392d29f0d9b53e2e70d",
"pubkey": "-----BEGIN PUBLIC KEY-----\nMCowBQYDK2VwAyEArDPmR+Gvu1HawnSvM6xjZao7BVPPQ80Qq3sa1hAYvDU=\n-----END PUBLIC KEY-----\n"
},
"verify": "Recompute the event chain h_i = SHA-256(h_{i-1} + canonical(event_i)) over `events` → must equal content_digest. Recompute the oracle-stream digest from `events` → must equal replay.cassette_digest. The signature covers `${content_digest}:${cassette_digest}`. Ed25519: verify with the public key embedded in manifest.signature.pubkey (no server key needed — fully offline-verifiable). HMAC-SHA256 (legacy): requires AUDIT_SIGNING_KEY. POST this record to /api/audit/verify.",
"spec_url": "https://runback.dev/spec",
"verifier_url": "https://runback.dev/verify"
},
"run": {
"run_id": "run_a3f1b90c",
"name": "loan-approval-agent",
"status": "error",
"input": "Applicant #4471 requests a $12,000 personal loan.",
"output": null,
"error": {
"name": "PolicyBlock",
"message": "issue_approval blocked — manual-review-required"
},
"metadata": {
"demo": true
},
"step_count": 2,
"total_tokens": 1030,
"started_at": "2026-06-14T02:47:11.000Z",
"ended_at": "2026-06-14T02:47:13.950Z",
"actor_type": null,
"actor_id": null
},
"events": [
{
"schema_version": 1,
"run_id": "run_a3f1b90c",
"span_id": "r",
"parent_span_id": null,
"seq": 0,
"ts_start": "2026-06-14T02:47:11.000Z",
"ts_end": null,
"type": "run",
"phase": "start",
"name": "loan-approval-agent",
"input": "Applicant #4471 requests a $12,000 personal loan.",
"output": null,
"status": "running",
"error": null,
"metadata": {
"demo": true
},
"_hash": "07bda052e5ce563160c1730ddefb62bace30cbeb871190097ee081d2a8570094"
},
{
"schema_version": 1,
"run_id": "run_a3f1b90c",
"span_id": "l1",
"parent_span_id": "r",
"seq": 1,
"ts_start": "2026-06-14T02:47:11.000Z",
"ts_end": "2026-06-14T02:47:12.400Z",
"type": "llm",
"model": {
"provider": "openai",
"model_id": "gpt-4o"
},
"request": {
"system": "You are a careful lending agent. Follow policy.",
"messages": [
{
"role": "user",
"content": "Applicant #4471 requests a $12,000 personal loan."
}
],
"tools": [
{
"name": "check_credit_file",
"description": "Pull the applicant's credit file.",
"parameters": {
"type": "object",
"properties": {
"applicant_id": {
"type": "number"
}
}
}
}
],
"params": {
"temperature": 0
}
},
"response": {
"text": "Checking the credit file before deciding.",
"reasoning": null,
"finish_reason": "tool-call",
"tool_calls": [
{
"tool_call_id": "t1",
"tool_name": "check_credit_file",
"input": {
"applicant_id": 4471
}
}
]
},
"usage": {
"input_tokens": 428,
"output_tokens": 184,
"total_tokens": 612
},
"latency_ms": 1400,
"error": null,
"_hash": "8f985ca54090e255926fb55830ecf4ab4dc4f9f32b598e04f50e34aaf09745fb"
},
{
"schema_version": 1,
"run_id": "run_a3f1b90c",
"span_id": "t1",
"parent_span_id": "l1",
"seq": 2,
"ts_start": "2026-06-14T02:47:12.400Z",
"ts_end": "2026-06-14T02:47:12.618Z",
"type": "tool",
"tool_name": "check_credit_file",
"tool_call_id": "t1",
"input": {
"applicant_id": 4471
},
"output": {
"score": 684,
"band": "near-prime",
"disputes_open": 1
},
"latency_ms": 218,
"error": null,
"_hash": "0fad198fd6469eb12aaa162d8acb5a0d9002f5eaa88806f4a4b6a79383b999bf"
},
{
"schema_version": 1,
"run_id": "run_a3f1b90c",
"span_id": "l2",
"parent_span_id": "r",
"seq": 3,
"ts_start": "2026-06-14T02:47:12.618Z",
"ts_end": "2026-06-14T02:47:13.900Z",
"type": "llm",
"model": {
"provider": "openai",
"model_id": "gpt-4o"
},
"request": {
"system": "You are a careful lending agent. Follow policy.",
"messages": [
{
"role": "user",
"content": "Applicant #4471 requests a $12,000 personal loan."
},
{
"role": "tool",
"content": "{\"score\":684,\"band\":\"near-prime\",\"disputes_open\":1}"
}
],
"tools": [
{
"name": "issue_approval",
"description": "Approve the loan.",
"parameters": {
"type": "object",
"properties": {
"amount": {
"type": "number"
}
}
}
}
],
"params": {
"temperature": 0
}
},
"response": {
"text": "Credit is sufficient — approving.",
"reasoning": null,
"finish_reason": "tool-call",
"tool_calls": [
{
"tool_call_id": "t2",
"tool_name": "issue_approval",
"input": {
"applicant_id": 4471,
"amount": 12000
}
}
]
},
"usage": {
"input_tokens": 286,
"output_tokens": 132,
"total_tokens": 418
},
"latency_ms": 1282,
"error": null,
"_hash": "76dbdec70aef43d145f8668c6401248be453f78655bba38957c33a36bf83cfe8"
},
{
"schema_version": 1,
"run_id": "run_a3f1b90c",
"span_id": "t2",
"parent_span_id": "l2",
"seq": 4,
"ts_start": "2026-06-14T02:47:13.900Z",
"ts_end": "2026-06-14T02:47:13.950Z",
"type": "tool",
"tool_name": "issue_approval",
"tool_call_id": "t2",
"input": {
"applicant_id": 4471,
"amount": 12000
},
"output": null,
"latency_ms": 50,
"error": {
"name": "PolicyBlock",
"message": "Blocked: an open dispute requires manual review before approval."
},
"policy_block": {
"rule": "manual-review-required",
"detail": "when credit_file.disputes_open gt 0 → must call escalate_to_human"
},
"policy_evaluated": {
"passed": false
},
"_hash": "bad224a034ed84cd8a216847e2f37084e5b7428f32b0868737faffcc007408a7"
},
{
"schema_version": 1,
"run_id": "run_a3f1b90c",
"span_id": "re",
"parent_span_id": null,
"seq": 5,
"ts_start": "2026-06-14T02:47:13.950Z",
"ts_end": "2026-06-14T02:47:13.950Z",
"type": "run",
"phase": "end",
"name": "loan-approval-agent",
"input": null,
"output": null,
"status": "error",
"error": {
"name": "PolicyBlock",
"message": "issue_approval blocked — manual-review-required"
},
"metadata": {},
"_hash": "2e91e3f5fc451d2c9841d9e3b6c68d78a9b2fca00fdff98989512704180a9189"
}
]
}Runback wraps the model calls your agents already make. Add it to one agent today; it's live from the next request. Setup is a developer task of about fifteen minutes; everything after it — replay, policies, audit export — is used from the dashboard.
import { withDebugger } from "@runback/sdk";
const dbg = withDebugger(model, { runName: "support-agent", redact: "standard" });
const res = await generateText({ model: dbg.model, tools: dbg.tools(myTools), prompt: task });
await dbg.finish({ output: res.text, status: "success" });When risk, audit or a regulator asks what your agent did six months ago, what do you hand them? A screenshot, a trace nobody can re-run, an export nobody can check — or a signed record they can replay and verify themselves.
An agent breaks in production and nobody can say why. Replay the decision, fix it — that incident becomes a permanent test in one click.
Get started free →An agent approves something a regulator later asks about. Every decision is sealed the moment it happens — one record covers APRA CPS 230, EU AI Act Art. 12, and NIST AI RMF. Self-hosted in your VPC.
Enterprise & compliance →Opposing counsel — or your own client — asks exactly what the AI said and why. Every session is captured, sealed, and replayable for court; high-risk outputs are flagged before they reach the lawyer.
Governance for legal →The decision that can't be wrong is the one you can't watch every time. A policy gate runs before the call — the block is sealed into the record.
How it works →Take the recording away and the other three are guesswork.
Every model call — context, tools, tokens — captured at the boundary. PII redacted in-process before anything leaves your app.
Capture →Re-run from the exact captured context — tools, retrieval, messages[] held fixed. A different output means behaviour changed. Root cause in minutes, not a war room.
Walk the demo →Simulate a policy against 90 days of decisions. Enforce it live — block the violating action before it runs.
Policy gates →SHA-256 hash-chained, Ed25519-signed record. Tamper-evident. The artifact your compliance team actually asks for.
Enterprise →Every incident becomes a regression test — why this compounds →
We're early. Rather than dress that up with stock-photo testimonials, we'd rather you judge this on the worked incident and the record you just verified above.
Walk the incident →Every layer of the replay engine is proven in public CI. We don't get to say "it works" without also showing you the run that proves it.
See the CI runs ↗Start free on managed cloud, clone the source and self-host, or book a call — we'll walk it through live in your environment.
Managed plans from A$49/mo — full pricing →