Execution lineage for AI agents
Execution lineage is the connected record of an agent run from intent through policy decisions, tool calls, human review, and resulting state. It preserves the identities, versions, sequence, and integrity references needed to replay and verify the execution.
KLA creates a Lineage Record for each governed run, correlates every decision and side effect, and anchors the record in an append-only evidence ledger for later review and export.
- 01IntentReview claim CLM-20418
- 02Governed execution3 tool calls · 2 policy decisions
- 03Human decisiondr_01K0A80D · approved
- 04Integrity recordsha256:9e8c…44a1 · sealed
- Primary record
- Lineage Record
- Investigation surface
- Lineage Explorer
- Integrity model
- Hashes, signatures, and ledger proofs
- Portable output
- Sealed Evidence Bundle
01: Concept
How execution lineage works in operation
Useful lineage preserves the causal path of the execution and the control state that governed each consequential step.
Execution lineage is the connected record of an agent run from intent through policy decisions, tool calls, human review, and resulting state. It preserves the identities, versions, sequence, and integrity references needed to replay and verify the execution.
- Start with stable identity
- The record binds the agent, Release, principal, tenant, environment, Process, and correlation identifiers for the complete run.
- Capture the actual sequence
- Inputs, model events, tool requests, policy decisions, approvals, responses, and state changes remain ordered and connected.
- Preserve control context
- Each governed action carries its authority snapshot, policy version, matched rules, outcome, and related Decision Request.
- Make integrity verifiable
- Hashes, signatures, append-only ledger anchors, and bundle manifests support tamper detection and Independent Verification.
02: KLA implementation
How KLA implements execution lineage
KLA builds the record during execution so operators and reviewers can inspect one continuous path from request to outcome.
- 01
Open the Lineage Record
A governed run receives stable execution, agent, Release, Process, tenant, and environment identifiers.
Output · Execution identity
- 02
Append runtime events
OpenTelemetry spans and KLA events capture inputs, model operations, tool calls, policy decisions, and responses in order.
Output · Ordered event path
- 03
Attach decisions and effects
Decision Requests, reviewer rationale, tool results, errors, and before and after state stay correlated to the relevant action.
Output · Complete operational record
- 04
Seal and export evidence
Integrity metadata anchors the record in the evidence ledger and supports export through the Evidence Room.
Output · Verifiable evidence
Example · Insurance claim
One record carries the claim from intent to effect
A claims agent gathers policy data, evaluates a proposed settlement, reaches a human checkpoint, and updates the claim system.
Lineage Explorer replays the exact sequence with tool inputs and outputs, policy and reviewer decisions, source references, and the final claim state.
- Execution startedreceived
claims-review-agent rel_2026_07_16_2 · claim CLM-20418
- Evidence retrievedrecorded
policy.read and documents.fetch · responses hashed and correlated
- Settlement held and approvedapproved
Policy threshold matched · claims supervisor rationale attached
- Claim state updatedrecorded
claims.update-settlement · before and after state hashes recorded
04: Evidence record
What KLA records for review
A Lineage Record connects technical events to the identities, control decisions, and business outcome that give the run meaning.
| Record layer | Captured evidence | Review purpose |
|---|---|---|
| Execution identity | Execution, correlation, agent, Release, Process, principal, tenant, and environment IDs | Define the governed run and its operating context |
| Event sequence | Inputs, prompts, model events, tool calls, responses, errors, and timestamps | Replay the path and preserve causal order |
| Control state | Authority snapshots, policy versions, verdicts, reason codes, Decision Requests, and rationale | Show how each consequential step was governed |
| Integrity and outcome | Before and after state, artifact hashes, signatures, ledger anchors, and bundle manifest | Verify the operational effect and detect later changes |
05: Connected controls
Follow the complete governed action path
The four concepts operate together across one action. Continue with the control layer closest to your next question.
06: FAQ
Questions about execution lineage
Definitions, runtime behavior, integration, and evidence boundaries for this control layer.
- What is execution lineage for an AI agent?
- Execution lineage is the connected, ordered record of an agent run. It includes the agent and Release, inputs, tool calls, policy decisions, human review, outputs, resulting state, and integrity references.
- How is Execution Lineage different from application logs?
- Application logs are usually service-specific operational messages. KLA Lineage Records correlate the full governed run across services and connect each consequential action to identity, authority, policy, human decisions, outcomes, and evidence integrity.
- Can a reviewer replay a Lineage Record?
- Yes. Lineage Explorer presents the ordered execution timeline with tool inputs and outputs, policy outcomes, Decision Requests, reviewer rationale, errors, and the resulting state.
- How does KLA make lineage tamper-evident?
- KLA hashes evidence artifacts, anchors records in an append-only ledger, and includes signatures and manifest hashes in Sealed Evidence Bundles. Independent verification can detect changes to the exported evidence.
- Can Execution Lineage be exported for an audit?
- Yes. Evidence Room packages selected Lineage Records, policy and approval evidence, integrity metadata, and a signed manifest into a Sealed Evidence Bundle or Control Pack.
Start with one action
Review the complete lineage of one agent action
Bring a consequential Process and its current logs. The KLA team will map the identities, decisions, effects, and integrity evidence a reviewer needs.