Product — Capture every AI agent action

§ IThe Capture Layer

Drop-in for the agent stack you already run.

Wrap OpenAI Agents SDK, Claude Agent SDK, MCP, LangChain, CrewAI, Vercel AI SDK, Mastra, LangGraph, or any bespoke agent in one line — in Python or TypeScript. Every tool call, model call, sub-agent spawn, decision and external action becomes a tamper-evident receipt.

Inspectable — your security team and your auditor can read the SDK end-to-end
Async non-blocking transport — never slows your agent
Local disk buffer for offline or air-gapped operation
Inline PII redaction — raw data never leaves your perimeter
Customer-held signing keys at the enterprise tier

Explore the SDK →

python · 5-line install

# pip install agentaudits

import agentaudit



agentaudit.init(

    api_key="aa_live_...",

    agent_id="claims_triage_v3",

)



# Your agent runs as normal —

# Agent Audit captures every action.

agent.run(customer_query)



# When the regulator asks, you have

# 1 click to an evidence pack.

receipt · v1 · evt_8f2a4b…

{

  "event_id":      "evt_8f2a4b...",

  "agent_id":      "claims_triage_v3",

  "ts":            "2026-06-07T14:23Z",

  "action": {

    "type":        "tool_call",

    "name":        "lookup_customer",

    "params_hash": "sha256:a1b2..."

  },

  "resource": {

    "classification": ["PII", "financial"]

  },

  "redacted_input":  "lookup [REDACTED]",

  "redacted_output": "{age: 34, ...}",

  "prev_hash":      "sha256:c4d2...",

  "signature":      "30450221..."

}

§ IIThe Receipt

A unit of evidence your auditor recognises.

Each receipt is a small, signed JSON object containing only what an auditor needs to understand the action. Sensitive payloads are fingerprinted, not stored — the hash proves the data existed; redaction proves you didn't keep it.

SHA-256 hash chain — retroactive edits are mathematically detectable
ECDSA signature — bound to your tenant's customer-held key
Schema v1, version-locked — never breaks downstream tooling
Optional RFC 3161 notarisation on every chain-head advance — independently timestamped, court-ready evidence
Binds into a signed as-of record — what authority was in force the instant the agent acted, pinned then, not reconstructed later

§ IIIThe Operator Surface

Every agent's life, queryable in seconds.

"Show me every action this agent took on this customer between these dates" is the primary query — and the one your existing logs can't answer.

№CapabilityWhat it does

01 Span-tree drill-down Each session renders as a tree of agent → sub-agent → tool calls, with input, output, decisions, classification, latency and cost inspectable on every node.

02 Filter by anything Agent, session, customer ID, data class, decision outcome, error, time range. All composable, all sub-second.

03 Verify integrity and completeness One click runs full hash-chain verification across any range of receipts — offline, confirming nothing was altered and nothing hidden. Catches tampering, omission and back-dating, with the first break index shown for forensic precision.

04 Anomaly alerts Decision drift, unauthorised tool calls, data-class escalations and spend spikes are surfaced automatically and pushed to Slack or Microsoft Teams.

05 SIEM integrations Webhook events into Splunk, Datadog, Microsoft Sentinel and Elastic — receipts live in your existing security telemetry pipeline.

06 UK data residency Default storage is London. Bring-your-own S3 and full self-hosted options at the enterprise tier — your data stays where your DPO says it should.

07 Occurrence vs authorisation Two separate verdicts on any action: that it happened (tamper-evidently) and that it was allowed (in force, in scope, delegated, not revoked). Shown side by side, never conflated — and never a claim of legal validity.

08 Delegation & scope checks Who authorised the authoriser, and was the action within scope — the delegation chain resolved as of the moment the agent acted, not reconstructed after the fact.

See the dashboard in detail →

§ IVThe Moat

The compliance answer your regulator already wants.

The moat isn't format coverage. It's the reproducible, independently verifiable as-of record under every pack — one signed, timestamped artefact that reproduces byte-for-byte: at the moment X happened, what was authorised, and what has happened since. Anyone can verify it; no one has to trust us.

On top of that, we maintain the export pack catalogue jurisdiction by jurisdiction. As regulations evolve, the packs evolve with them — no spreadsheet rewrites, no Excel macros, no consultant rebuilds.

№ 01 EU AI Act Article 12 High-risk system record-keeping evidence covering every operational obligation under Article 12. Available

№ 02 ICO Subject Access Every action involving a named data subject + lawful basis. Article 22A–22D ADM disclosures included. Available

№ 03 FCA SYSC Operational resilience evidence demonstrating AI-driven service continuity for FCA-regulated UK firms. Available

№ 04 NIST AI RMF Govern / Map / Measure / Manage evidence pack for US enterprise diligence. Available

№ 05 Insurance claim pack Cyber and E&O policy claim evidence — incident timeline + affected receipts. Available

№ 06 Board / Audit Committee Quarterly governance summary, anomaly review, material decisions logged. Available

See every pack →

One capture layer.
Every agent.
Every regulation.

Drop-in for the agent stack you already run.

A unit of evidence your auditor recognises.

Every agent's life, queryable in seconds.

The compliance answer your regulator already wants.

See it run against your stack.

One capture layer. Every agent. Every regulation.

Drop-in for the agent stack you already run.

A unit of evidence your auditor recognises.

Every agent's life, queryable in seconds.

The compliance answer your regulator already wants.

See it run against your stack.

One capture layer.
Every agent.
Every regulation.