A loan-approval agent ignores its own $50k limit and calls issue_approval for $85,000. With Runback: policy gate fires before the call executes, root cause is open in 2 minutes, the block is sealed into a tamper-evident cassette. Without Runback: 2,400 log lines, no context, 3 hours and 48 minutes of guessing.
Runback opens every failed run on the failure — no hunting. Click agent decidesto see the messages[] array the model received. That's where the root cause is.
Time from incident to root cause identified: 2 minutes. The logs gave you 2,400 lines and no context. The Context tab gives you the exact prompt.
AI agent decisions have had no equivalent of the git commit — immutable, attributed, tamper-evident. The Runback cassette is that primitive: append-only, hash-chained, sealed at execution. Not a log of what happened — the decision, preserved for replay. This is the cassette for run_a3f1b90c. Each entry extends the chain; add, remove, or alter one byte and verification fails. Why this matters for long-term proof →
| seq | kind | key | hash |
|---|---|---|---|
| 0 | now | now | 8f2a… |
| 1 | llm | llm:gpt-4o:a3f8… | c91d… |
| 2 | tool | tool:check_credit:c29a… | e7b3… |
| 3 | llm | llm:gpt-4o:b7d1… | 2a8f… |
| 4 | tool | tool:issue_approval:BLOCK← policy_block: { rule: "loan_limit", detail: "85000 > 50000" } | 09c4… |
This tool call never executed. The block itself is hash-chained — seq: 4, rule: loan_limit, detail: '85000 > 50000' is part of the sealed record. Replay can prove the agent was stopped, not just that it failed.
Policy blocks are first-class cassette entries with their own seq and hash. The policy predicate is purely deterministic — no LLM judge, no probabilistic scoring. If the rule fires, the entry is sealed and the call is blocked.
After root-cause and fix, the engineer clicks Save as golden testin the debugger. The run — this exact run, with this exact context — becomes a permanent entry in the regression suite. The next time the same class of failure occurs, the golden suite catches it in CI before it reaches production. The suite builds itself from your actual failures, not from someone's memory of what to test.
Start with the free Community edition — one agent, your own VPC, no data leaves your network. The first incident tells the story.