Blog · ← All posts
The Five Types of AI Agent Records: Actions, Decisions, Outcomes, Refusals and System Events
By the Ingenious AI team · Published: November 14, 2026 · Reviewed: November 14, 2026 · 6 min read
The short answer: A complete agent ledger interleaves five types of records: actions (what the agent did — the tool calls), decisions (why — the stated reasoning, policy version and approvals), outcomes (what came back — results, downstream IDs, errors), refusals (what the agent or its guardrails declined — and what happened next), and system events (the lifecycle facts — enrollment, state changes, pauses, verification runs). Each type answers a different audience question, and a ledger missing any one of them has a characteristic blind spot that incidents and audits find on schedule.
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.
Most agent logging debates treat 'the log' as one undifferentiated substance — more of it or less of it, kept longer or shorter. The teams that get value from their records think in types, because different records do different evidentiary work: an action record proves an event, a decision record proves a basis, a refusal record proves a boundary. Knowing the five types turns logging design from a volume decision into a coverage checklist.
This guide defines the five types as they appear in Agent Flight Recorder's ledger, states what each one proves and what breaks when it is missing, and shows how they interleave into the single hash-chained sequence investigators actually read.
What is a record type? a category of ledger entry distinguished by what it attests: actions attest execution, decisions attest basis, outcomes attest results in the world, refusals attest boundaries encountered and respected (or tested), and system events attest the agent's own lifecycle and administration. Typing records at capture time — rather than inferring types later from content — is what makes interrogation, anomaly rules, and audit sampling precise.
Type 1: Actions — the verbs of the history
Action records are the spine: every tool call, with its masked argument summary, digest, timestamp, and the agent's identity. They answer the first question of every inquiry — what did it do? — and they are the unit that downstream reconciliation matches against the world's own books. A ledger of only actions is a silent film: everything visible, nothing explained. It is the commonest real-world ledger, because actions are the easiest type to capture, and its characteristic failure arrives at the first 'why' an auditor asks.
Type 2: Decisions — the basis, on the record
Decision records capture the why at the moment it existed: the agent's stated reasoning in a sentence or two, the policy or prompt version in force, the plan where one formed, and any approval attached. They are the smallest records in the ledger and the highest in value per byte, because they convert a sequence of acts into an account of intent that can be judged — against the mandate, against the policy, against what a reasonable operator would have wanted. Missing decisions, every investigation becomes an exercise in reverse-engineering intent from behavior, and every dispute becomes a contest of interpretations the records were supposed to settle.
Type 3: Outcomes — closing the loop with the world
Outcome records state what the tool returned: success with the downstream identifier, failure with the error class, or partial completion flagged as partial. They look redundant — until the lost-acknowledgement incident, the duplicate-action dispute, or the auditor's trace from the processor's side arrives, at which point the outcome record is the one that agrees with the external witness. Outcomes also feed the honest metrics: error rates, retry patterns, and the slow degradation of a tool contract that action records alone would render as unexplained changes in agent behavior.
Type 4: Refusals — the boundaries, proven working
Refusal records capture declinations in both directions: the guardrail or tool that refused the agent (cap exceeded, allowlist miss, approval absent), and the agent's own refusals of user or injected instructions. Each refusal record includes what the agent did next, because the sequel is the story — a refusal followed by compliance is a control working; a refusal followed by a creative workaround is the opening scene of an incident, visible in two adjacent records. Teams that log only successes delete this type by default and with it their best evidence that boundaries exist, plus their earliest warning when an agent starts testing them.
Type 5: System events — the agent's own biography
System events record the administration of the agent itself: enrollment, mandate changes, promotions, suspensions and resumptions, pauses during incidents, verification runs and their results, certificate mintings, retirement. They answer the authority questions — should this agent have been running at all that day, under whose mandate, and who changed it — and they give the other four types their interpretive frame: an action judged against the mandate in force requires knowing when mandates changed, which is a system event's testimony. Omitting this type leaves the ledger describing a agent that was never born, never governed, and never stopped — a story no auditor believes, because no real system lives that way.
Five actions you must take today
Audit your ledger against the taxonomy — each type gets one test:
- Actions: pick any downstream event from last week and find its record in under a minute. If you cannot, coverage — not typing — is your first problem.
- Decisions: open a consequential action and read its stated basis and policy version from the record itself. If the basis lives in a chat transcript elsewhere, the type is missing.
- Outcomes: find a failed tool call from the records alone, with its error class and what the agent did next. Success-only logging fails this test silently.
- Refusals: produce the last five guardrail refusals and their sequels. If the answer requires interviewing engineers, the boundary has no evidence.
- System events: state, from the record, the mandate in force on a date three months ago and who approved it. If that requires a wiki archaeology session, the biography is unwritten.
Worked example: one run, five types, complete story
A single support run at fictional retailer Fern & Field shows the interleaving. It opens with a system context (the agent active under mandate v6). A decision record states the plan: verify the order, then refund within cap if damage is confirmed. Action records follow — order lookup, photo review — each with its outcome record (found; image classified as transit damage). A refusal appears mid-run: the agent attempted a goodwill credit above its cap, the tool refused, and the next decision record shows it proceeding with the in-cap refund and a note to the customer — the boundary, the respect for it, and the recovery, in three records.
The refund's action and outcome records close the run, the outcome carrying the processor's ID. Total: 23 records, five types, and a run that an auditor, a dispute reviewer, or a new engineer can read cold and understand completely — what was done, why, what came back, where the line was, and under whose authority the actor operated. Remove any one type and a specific, predictable question becomes unanswerable; that is the taxonomy's whole argument.
FAQ
Do decisions and actions need to be separate records?
They need to be separately retrievable facts; whether they ride as one combined record or two adjacent ones is an implementation choice. AFR's practice — the reasoning carried with the action record, typed and queryable — works because interrogation can address the decision layer independently ('show every action whose reasoning cites an external document'). What fails is reasoning kept in a different system from actions, where the join depends on timestamps and optimism.
Are system events part of the hash chain or separate?
Part of the chain. Administration is exactly where a motivated editor would work — backdating a mandate change, erasing a suspension — so lifecycle records get the same tamper-evidence as actions, in the same sequence. A separate, mutable admin log next to a chained action ledger is a door labeled 'the part you may edit', and reviewers will price the whole system by it.
What about records of human actions in the workflow?
They belong, typed as what they are: approvals, instructions, overrides, pauses — entered as events in the agent's stream or as cause records the agent's actions cite. The five-type taxonomy extends naturally (an approval is a decision record with a human author), and the principle is uniform: anything that explains the agent's behavior belongs in the same verifiable history as the behavior.
How do the types affect retention?
Mostly uniformly — all five are small and belong to the long tier — with one refinement: refusal and system-event records often outlive even the action records' usefulness, because governance history (when mandates changed, how boundaries behaved over years) is the material of trend audits and culture assessments. Storage being trivial, the practical rule is to retain all five on the longest clock rather than engineering distinctions nobody will thank you for.
Can we retrofit types onto an existing log?
Partially, going forward: start typing at capture now (actions and outcomes usually exist implicitly and can be typed by instrumentation; decisions require the agent to emit them; refusals require the guardrail layer to speak up; system events require the admin path to write records). The historical, untyped period remains readable but coarser — one more entry in the ledger of reasons to start the complete design before the history that matters accrues.
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.