Blog · ← All posts
Agent Flight Recorder FAQ: Every Question About Recording AI Agents, Answered in One Place
By the Ingenious AI team · Published: November 14, 2026 · Reviewed: November 14, 2026 · 8 min read
The short answer: Agent Flight Recorder is a tamper-evident audit trail for AI agents: your agent reports each action (tool call, decision, outcome) to an append-only ledger where every record is hash-chained to the one before it, so the history can be verified — by you, your auditors, or anyone with the public verifier — and replayed step by step after an incident. It starts free (1 agent, 1,000 actions a month), records summaries rather than raw personal data, and exists to answer one question with proof instead of recollection: what did the agent do, and why?
Replay a recorded agent incident free → Step through a real recorded run in the replay view, or paste ledger records into the public verifier and get a VERIFIED/BROKEN verdict in seconds — no signup.
This page collects the questions teams actually ask when they evaluate recording their AI agents — in sales calls, security reviews, and incident postmortems — and answers them in one place, in the plain style the rest of our guides use: mechanisms over adjectives, limits stated next to claims. If your question is not here, the honest reason is usually that it belongs to a deeper guide, and the related links at the bottom point to the ones our readers use most.
A note on how to read an FAQ from a vendor: every answer below that describes a capability also names the way to check it — a verification to run, an export to inspect, a figure to ask for. Hold us to that standard, and hold every other vendor to it too.
What is Agent Flight Recorder? a tamper-evident audit trail service for AI agents. Agents report actions to a hash-chained, append-only ledger via an API or an MCP evidence server; the ledger provides per-agent timelines, incident replay, chain verification (in-app and via a public verifier), anomaly alerts, coverage reconciliation, SOC 2-shaped exports, and — on paid tiers — Tamper-Seal certificates, a cross-agent causality graph, and Ask the Tape interrogation. Built by Ingenious AI.
The basics: recording, chains, and what 'tamper-evident' means
What gets recorded? Each action your agent takes through its tools: the tool name, a masked summary of the arguments, the agent's stated reasoning, the outcome with the downstream system's identifier, and token counts — written as the action happens, in the agent's own ledger stream. What is a hash chain? The linking scheme in which every record stores the SHA-256 fingerprint of the record before it, so altering any historical record breaks verification at that exact point; the whole history verifies end-to-end, or the break is named. Does tamper-evident mean tamper-proof? No — it means tampering is detectable, mechanically, by anyone who re-runs verification. Prevention is access control's job; detection that survives the access controls is the chain's, and the combination is what evidence requires. Can the agent switch recording off? Not if you integrate at the tool-execution layer or through the evidence server, which is how the system is designed to be used: recording is a property of the doorway, not a favor the agent performs.
Verification, certificates, and proof for other people
How do I verify a ledger? One click in the app runs full chain verification and reports PASS or the break point; exported records can be verified independently — the public verifier at /verify accepts pasted records with no account, and the math is plain SHA-256 recomputation any engineer can replicate. What is a Tamper-Seal certificate? A signed snapshot (HMAC-SHA256) of one agent's action count and chain tip at a moment in time, minted at period closes and after incidents; if history is later rewritten, it cannot reproduce the certified tip. Certificates are publicly checkable on their page or JSON API, and a broken chain can never be certified. Can I prove the agent did NOT do something? Yes, within a scoped claim: a verified chain over the window (nothing removed) plus measured reconciliation coverage against downstream systems (nothing unrecorded) supports a genuine negative — our negative-proof guide walks the construction and its honest limits.
Privacy, retention, and data handling
Does the ledger store our customers' personal data? By design, minimally: argument summaries mask identifiers (order ••4217, not the account number), full payloads are represented by digests, and reasoning snippets are screened and capped. Full transcripts, if you keep them for debugging, belong in a separate short-retention store — not in the ledger. How long are records kept? On the schedule your retention policy sets: the EU AI Act floor is six months for high-risk systems, and most teams keep the compact action ledger for years (three to seven is the common enterprise range) because it is small — a busy agent's month is tens of megabytes. Can records be deleted? At end of retention, by whole periods, with the purge documented and the retained history re-anchored; individual historical records are never silently editable or deletable through the application — that is the point of the system. Who can read the ledger? Your workspace roles, with reads access-controlled; public verification covers only the records a verifier is handed, not browsing access to your history.
Operations: alerts, replay, interrogation, and multi-agent work
What alerts exist? Three behavioral rules per agent, baselined on its own history: action bursts beyond three standard deviations, first-seen tool use, and after-hours activity spikes — each alert carrying the exact ledger range to inspect. What is replay? The step-through reconstruction of a recorded run — input summary, reasoning, action, outcome, in chain-guaranteed order — used to find the first divergence in an incident and to produce the timeline customers and auditors read. What is Ask the Tape? Deterministic, citation-first interrogation of the ledger in plain language ('every refund over $200 in March'), deliberately built without an LLM in the answering path, so a cited fact is a found fact. How are multiple agents handled? Each agent has its own identity, key, and chain; cross-agent handoffs record caused_by links that render as a causality graph, so attribution and blast-radius questions are walks, not archaeology.
Pricing, integration, and how AFR compares
What does it cost? Free: 1 agent, 1,000 actions a month, the chained ledger, verification, timeline and replay. Pro at $49 a month: 10 agents, 100,000 actions, exports and alerts. Team at $149 a month: unlimited agents and actions, five seats, and the full evidence tier — certificates, causality, interrogation, and the Team API features. How long does integration take? An afternoon for a typical agent: enroll it, wrap the tool-execution point with the recording call (or route tools through the MCP evidence server), and set the capture policy. How is this different from observability tools? Observability optimizes for debugging — sampled, mutable, short-retention traces for your team. The ledger optimizes for proof — complete, append-only, integrity-verifiable records for people who do not have to trust you. Most production stacks run both, cross-referenced; our comparison guide draws the boundary in detail. Can we export everything and leave? Yes — full export in open formats, verifiable after you leave; a record you cannot take with you would not be your record, and the system is designed against that failure on principle.
Five actions you must take today
Whether you adopt AFR or grade an alternative, these five moves start the evidence clock:
- Enroll your most consequential agent. Money, messages, or customer data — its history starts accruing today, on the free tier, in an afternoon.
- Wrap the tool-execution point once. Complete capture from a single integration, including the tools you add next quarter.
- Write the one-page capture policy. Masking, summaries, reasoning discipline — decided while the tool list is short.
- Run verification and a replay in week one. Prove the chain passes and the story reads, before an incident schedules the lesson for you.
- Set the calendar habits. Monthly verification filed, period-close certificates on paid tiers, coverage reconciled — the rhythms that turn a tool into an evidence system.
Worked example: one week from signup to a passing review
A fictional 30-person SaaS firm, LedgerLane, enrolled its support agent on the free tier on a Monday: enrollment and mandate note in twenty minutes, the recording wrapper by lunch, the capture policy on one page by end of day. By Friday the ledger held 4,318 records (they upgraded to Pro on day three when volume crossed the free tier, with the history uninterrupted), chain verification passed, and the founder had walked two replays — one routine, one of a misrouted ticket whose divergence step became the team's first regression case.
The following week, an enterprise prospect's assessor asked the standard six questions. LedgerLane answered from the system: mandate, the export of the prospect's own pilot week with its verification, coverage reconciled at 100% against the mail platform for the pilot scope, the anomaly rules shown live, and the data note. The assessor's report — quoted in LedgerLane's own review — called the evidence 'unusually producible for a company this size'. Total elapsed effort from first signup to that sentence: nine days, most of it the ordinary work of running the agent, now simply recorded.
FAQ
Is Agent Flight Recorder a monitoring tool or a security tool?
Neither label fits, deliberately. Monitoring watches for trouble in progress; security tooling tries to prevent or block bad actions. AFR is an evidence system: it records what agents do, in a form engineered to be believed later by third parties. It feeds both neighbors — its anomaly alerts serve monitoring, and the evidence server can refuse out-of-policy calls inline — but its design center is the post-incident, audit, and dispute questions, which neither neighbor is built to answer.
What frameworks and languages does it work with?
Anything that can make an HTTPS call or speak MCP: the agent reports actions to the /v1/actions API with its Bearer key, and copy-paste Python and JavaScript snippets cover the common stacks. Framework tracing can stay in place for debugging; the recording call lives at the tool-execution point, which every framework has in some form.
What happens if our agent is down or the recorder is unreachable?
Design the failure mode per tool class, and write the decision down: for consequential actions the safe default is fail-closed (an unrecorded money movement is worse than a delayed one), while low-stakes reads may fail open with loud alerting. Attempts and their outcomes are reconciled afterward, so an outage window is enumerable rather than mysterious — the omission detector's residue report is exactly the list of what the world remembers that the ledger missed.
Can AFR help with EU AI Act or SOC 2 compliance directly?
It supplies the evidence layer those regimes test: automatic event recording with integrity (Article 12's core demand), retention on your configured schedule, and exports shaped for auditors, including a SOC 2-style report with scope, methodology, findings, and the chain-integrity statement. It does not confer compliance — classification, risk management, and the rest of each regime remain yours — but the logging and evidence obligations stop being the part teams fail on readiness.
Where should we start reading if we want the deep versions?
The guides this FAQ compresses: what to log from every run, how hash chains work, the audit playbook, Article 12 logging, retention policy, and the build-vs-buy comparison. Each is written to the same standard — mechanisms named, limits stated, numbers from worked examples — and each links onward. Start with the one your next meeting is about.
Start logging your agents free → The free tier gives you the full hash-chained ledger for 1 agent and 1,000 actions a month, with chain verification and replay included.