# How to audit your claw — see exactly what your AI agent did **URL:** \ **Category:** Support & FAQ\ **Created:** [September 9, 2026, 3:16pm UTC](https://community.sqs.chat/t/how-to-audit-your-claw-see-exactly-what-your-ai-agent-did/390 "2026-09-09T15:16:44Z")\ **Posts on this page:** 1\ **Page:** 1
**Author:** ![tess\_walker](https://community.sqs.chat/letter_avatar_proxy/v4/letter/t/73ab20/32.png) [@tess\_walker](https://community.sqs.chat/u/tess_walker)\ **Post date:** [September 9, 2026, 3:16pm UTC](https://community.sqs.chat/t/how-to-audit-your-claw-see-exactly-what-your-ai-agent-did/390/1 "2026-09-09T15:16:44Z")
## How to audit your claw — see exactly what your AI agent did One question we hear a lot: _“Can I actually see what my agent did — with which model, on whose say-so, and whether anyone changed the record afterward?”_ Short answer: yes. Every claw keeps a tamper-evident record of the AI calls and decisions made on it, and you can ask for a verifiable copy at any time. Here’s what that means in plain terms and how to use it. ### What gets recorded When your agent runs, each significant step is written to a log that includes: - **What was asked** (a short, redacted preview of the prompt — not your full data) - **Which model answered** - **Whether it succeeded** and a preview of what it produced - **Who/what triggered it** and roughly when Decisions that matter — like an approval, a waiver, or a sign-off — are recorded the same way, so “who approved this and why” is on the record too. ### The part that makes it trustworthy: it’s tamper-evident The log isn’t just a file someone could quietly edit. Every entry is cryptographically chained to the one before it (think of a fingerprint that depends on the whole prior history). Two things make it hard to fake: 1. **Nothing can be silently edited or deleted** — the system rejects changes to past entries. 2. **The chain can be re-checked** — a verify step replays the whole history and will flag anything that was modified, removed, or reordered. So when we hand you an export, it isn’t “trust us, this is what happened” — it’s “here’s the record, and here’s how you can check nothing was changed after it was written.” ### How to get your audit The cleanest path is to ask us (support) for an export for your claw. To make the request: 1. Tell us **which claw** and **the date range** you care about. 2. We run a **chain verification** first — if it comes back clean, that’s the green light. 3. We hand you a **machine-readable export** of that window. You can also open a request right from your claw’s dashboard if you’d rather not email. ### How to read the export Each line is one event. The useful fields are: - **event type** — what happened (an AI request, a response, an approval, etc.) - **timestamp + tenant** — when and on which claw - **model** — which model was used - **prev\_hash / curr\_hash** — the chain links. You don’t need to compute these yourself; the verify step does that. If someone ever hands you a log that fails verification, that’s a red flag — tell us immediately. ### What it does _not_ mean A tamper-evident log proves the record wasn’t changed _after_ it was written. It doesn’t claim the agent was perfect — only that what it did is on the record and can’t be quietly rewritten. Garbage in, verifiable garbage out — which is exactly why you want both a good record _and_ the ability to watch what your agent is doing day to day. ### Want to go deeper? If you’re technically inclined, the same verification can be run on your export independently — the chain math is deterministic and documented. Ask support for the technical details or point your own engineer at the export; it’s designed so a third party can re-verify it without trusting us. * * * **Bottom line:** your claw keeps a tamper-evident record of what your AI agent did, and you can get a verifiable copy whenever you need it — for a client, for compliance, or just for peace of mind. Ask us for an export and we’ll verify + hand it over. For reference on why verifiable records matter in AI operations, see this overview of AI-agent auditability and governance ([Anthropic on AI agent observability and control](https://www.anthropic.com/engineering/building-effective-agents)). * * * _by Tess Walker — the help desk, Side Quest Studios_ _AI-assisted, curated for Side Quest Studios._ ## References 1. Anthropic on AI agent observability and control — [Building Effective AI Agents \ Anthropic](https://www.anthropic.com/engineering/building-effective-agents) _See also: [the deeper audit walkthrough](https://community.sqs.chat/t/444)._