Is this a startup problem or an enterprise problem?

Both — the same missing primitive, different stakes.

Seed-stage team

The missing primitive costs an engineer a bad night: "why did the agent do that at 2 AM" with no way to answer it except staring at logs.

Regulated enterprise

The exact same gap — no re-executable record of what the model saw and decided — becomes an unanswerable question from an auditor, a board, or a court.

Four failure classes show up at every scale. None are visible in a log. All are visible in a replay.

Policy violation

The rule existed, the model ignored it.

Context contamination

Retrieval skewed the decision.

Prompt injection

User input overrode the system prompt.

Model drift

A silent upstream update shifted behaviour.

End to end

Three ways this shows up in production — click through each one.

Live demos, not screenshots — we're early enough to show the mechanism rather than dress up a case study we don't have yet.

Incident → root cause

An agent does something wrong and nobody can say why.

loan-approval-agent auto-approves a loan it should have escalated. The only record is a log line: "approved, 2:47 AM." Below is an illustrative before/after from the incident walkthrough — a worked example, not a customer story. We are too early to have one.

Without Runback

3h 48m

Grep through 2,400 log lines.
Guess. Patch. Hope it doesn't recur.

›2:47 AM — alert fires

›3:00 AM — first engineer online

›4:15 AM — found a suspicious log line

›5:20 AM — maybe found the prompt issue?

›6:35 AM — fix deployed, fingers crossed

→

same incident

With Runback

4m 23s

Open the run. Root cause is highlighted.
Patch, replay to confirm, seal the record.

✓2:47 AM — alert fires with run link

✓2:49 AM — run open, failure step highlighted

✓2:51 AM — root cause found in messages[]

✓2:52 AM — fix applied, replay confirms

✓2:52 AM — sealed. Done.

Walk this exact scenario, step by step →

Prevention → policy gate

A rule everyone agreed on gets violated in production anyway.

A limit written into the system prompt is advisory, not enforced — the model can decide to break it anyway. Pick the world closest to yours, then click through the run: the gate catches it before the call reaches anyone.

support-agent · refund · blocked at gateclick any step ↓

Policy gate · issue_refund — blocked

amount: $250
rule: no_refund_over_100

blocked call

issue_refund({ amount: 250, customer_id: 8842 })

Gate blocked this: a $250 refund on a disputed charge breaks two policies. Caught before it reached the customer.

Change management → model drift

You want to upgrade the model. You don't know what it'll break.

The same captured context from the run above, replayed against two other models before anyone ships an upgrade. Same input, same tools — different output means behaviour changed.

support-agent · refund · replayed against 3 candidate models

claude-sonnet-4.5 (current)→ escalate(dispute)baseline

gpt-4o→ escalate(dispute)match

llama-3.3→ issue_refund(250)regression

Runback's behavioural drift view: a stability score for the last 7 days against a 30-day baseline, with agents monitored, unacknowledged alerts and alerts over the last 30 days.

See replay in the mechanism guide →

By where you are

The same four capabilities, applied to your actual stakes.

Early-stage AI team

Week 1: wrap your riskiest agent, replay your first real failure instead of guessing at it. By month one, every incident auto-mines into a regression test, so the same bug can't silently come back after a prompt change.

See the rollout →

Banks & financial institutions

Every decision sealed in a tamper-evident chain, self-hosted in your own VPC. APRA CPS 230, EU AI Act Art. 12, NIST AI RMF — one record, reviewed by your risk committee, not three separate log exports.

Enterprise & compliance →

Law firms & professional services

Every AI session captured and sealed the moment it happens. Replay the exact session for a partner, a court, or a regulator. Policy gates flag a high-risk output before it reaches a client.

Governance for legal →

Safety-critical deployments

Policy gates run before the call, not after the incident report. The block is sealed into the record automatically — so "we prevented it" is something you can show, not something you have to be believed about.

How it works →

What problem does this actually solve — end to end?