Run record
What triggered it, the batch, counts of completed, escalated and failed items, duration.
The short answer
A good trail lets anyone reconstruct each step: what the agent saw, what it did, which rule or model output decided it, what changed in the system, who approved, and when. Capture it per step, protect it from edits, keep secrets out, and keep it as long as the business record it supports.
Why a log line is not proof
An auditor asks four things: what happened, who did it, on what basis, and can you prove it. "Invoice 44817 processed" answers none of them. A chat transcript records what a model said, not what the system did.
- What?missingWhat it did
- Who?missingWho
- Why?missingWhy it decided
- Proof?missingSeal
Record every step
The step is the right unit. A run can have forty steps, and the one that went wrong needs its own record.
| Field | What to capture |
|---|---|
| Step identity | Workflow and version, run ID, step number |
| Time and place | Timestamp with time zone, machine, session |
| What it saw | The screenshot or page it read, and the region used |
| Values read | Each value and where it sat in the source |
| Why it decided | Rule name and version, or model output and what routed it |
| What it did | Target control, typed text (masked if sensitive) |
| What changed | The screen after, plus the system's response |
| Outcome | Success, failure, retries, escalation |
| Who | The agent, the user session, any approver |
| Seal | A hash of this record and the one before |
Three more records
Steps alone are not enough.
Change record
Every edit to a workflow, rule, table or model, with who, when, old value and new value. A wrong number next quarter usually traces to a change made this quarter.
Review record
Who reviewed, what they saw, their decision, any correction, the reason code, how long it waited.
What never goes in the log
- Passwords, session IDs, tokens and keys
- One-time codes, even though they flash on screen
- Bank, card and sensitive personal data in screenshots (mask at capture, not later)
- A model's free-form reasoning text (store the output, the decision and a short reason)
Proving nobody edited it
A log in a shared folder is a document anyone could change. Hash chaining fixes that cheaply. Each record stores a fingerprint of itself plus the fingerprint of the record before it. Change one screenshot and the chain breaks from that point on.
Export the evidence in a format an auditor can open without the vendor's software, and keep a copy operators cannot alter.
How long to keep it
The trail inherits the retention of the record it supports. Use the longest rule that applies: tax rules, contract audit clauses, regulators, or your own records policy.
Tier storage to keep it affordable: full screenshots for a recent window, then extracted values and hashes. Keep full images indefinitely for anything that failed or went to review.
How LaunchAI handles it
Every action, decision and run in LaunchAI is logged with screenshots, kept on the machine that did the work, and exportable in an audit-friendly format. Credentials and one-time codes stay out of the record. Product →
Questions
Is a screenshot trail sensitive data itself?
Yes. It holds whatever the agent saw. Encrypt it, restrict viewing by role, log who opens it, and apply the same retention as the source records.
How is this different from the ERP's change log?
The ERP log shows that a field changed and under which user. It cannot show what the agent was looking at, which rule decided, or who approved outside the ERP. An investigation needs both.
Keep learning
- Human in the Loop Put a person where a wrong action is expensive and hard to undo, and where the agent's evidence is weak. Read the guide
- Stop Hallucinated Numbers Never let the model have the last word on a number. Read the guide
- Testing Before Production Test an agent like a new hire with write access, only more often. Read the guide
No prompts. No babysitting. No Integrations.