0% read on this device
Browse the curriculum

Start Here

Vision

First Principles

Operating Model

Domain Model

Agent Factory

Runtime Architecture

AI Engineering

Autonomous Workflows

Verification & Delivery

Factory Platform

Quality Engineering

Security & Governance

Case Studies

Labs

Interview Practice

Research Journal

Reference

Curriculum/Runtime Architecture/A rapid review of the chapter’s existing Quick Read, principles, definitions, lessons, and review material.
Runtime Architecture5 min readchapter

Factory Observability and Agent Runtime Telemetry

An agentic factory can be busy while making no progress, cheap while producing unsafe work, or successful according to a model while violating authority. Operators need to understand outcome, control state, execution health, cost, and evide

Status: Draft for studyRisk: highLifecycle: executeContent reviewed 2026-08-11Maturity guide →
Claim boundaryThis is curriculum guidance. It does not by itself prove a production implementation.
study mode

A rapid review of the chapter’s existing Quick Read, principles, definitions, lessons, and review material.

Whiteboard exercise

Reconstruct and defend this chapter’s architecture.

Reconstruct the architecture, name each boundary, and defend the tradeoffs.

runtimeharnessenvironment5 min chapter
Open the source exercise

Trace one WorkOrder from dispatch through a timed-out Attempt, lease expiry, retry, PR publication, validation, and acceptance. Show span links, authoritative records, metrics, evidence, and alerts. Explain which data can be sampled and which cannot.

3. Enduring Principle

Observe from intent to outcome

Use a stable correlation spine:

Mission -> Plan revision -> WorkOrder revision -> Task -> Attempt -> trace -> span/tool call -> artifact/evidence -> PR -> release

Every signal should carry tenant/workspace/repository scope, relevant domain IDs, execution-manifest digest, source/head SHA when known, actor or workload identity, and timestamp. Correlation must not grant access: authorization still filters what each viewer may retrieve.

Separate authority, evidence, audit, and telemetry

RecordPrimary purposeMay be sampled?Controls advancement?
Domain stateCurrent authoritative workflow decisionNoYes
Audit eventWho decided or changed authorityNoSupports accountability
Evidence receiptProof for a claim about an exact subjectNoYes, through policy
Trace/log/metricDiagnose behavior and measure healthOftenNo, unless promoted through an evidence-producing control

A trace showing status=success cannot accept a WorkOrder. A validator can transform an attributable observation into an evidence receipt under an explicit method and policy.

Instrument the control and execution planes

Control-plane spans should cover Mission/Plan decisions, policy evaluation, WorkOrder release, dispatch, acceptance, and exception handling. Execution-plane spans should cover claim/renew/reconcile lease, sandbox creation, model invocation, tool call, command, test, commit, publication, receipt ingestion, and cleanup.

Use links rather than false parent-child relationships when asynchronous retries or external CI runs have independent lifecycles. Preserve attempt number and idempotency key so duplicate events can be explained.

Measure four kinds of health

Outcome: lead time to validated value, throughput, change failure, rework, acceptance, and customer signal.

Control: blocked gates, policy denials, exception age, approval latency, stale/missing/conflicting evidence, unauthorized attempts, and autonomy demotions.

Runtime: queue age, lease expiry, heartbeat lag, retry rate, timeout, cancellation latency, sandbox failures, publication failures, and reconciliation backlog.

AI economics and quality: model/provider, input/output/cache tokens, latency, cost, tool-call success, task/eval success, human override, and routing decision outcome.

Token volume and agent activity are diagnostic inputs, not productivity measures.

Capture decisions without exposing hidden reasoning

Record the decision input class, selected policy/routing rule, model and configuration, structured output, confidence where calibrated, tool choices, and result. Do not require or persist private chain-of-thought. Retain user content, prompts, tool arguments, and outputs only under classification, redaction, access, and retention rules. Default observability should be metadata-first and content-off.

Make freshness and silence observable

Evidence has valid-from/valid-until semantics and a subject digest. Alert when required evidence approaches expiry, the subject changes, a verifier becomes unhealthy, or no expected heartbeat/result arrives. Absence is a signal only when the system defines what should have happened and by when.

Design operator alerts around action

An alert should say what outcome is affected, why it matters, which record is authoritative, current risk, recommended action, owner, and deadline. Deduplicate by governed incident, suppress downstream symptoms, and escalate by consequence and duration. The factory UI should prioritize exceptions and decisions over a wall of agent activity.

Use open telemetry carefully

OpenTelemetry provides standard traces, metrics, logs, baggage, and context propagation. W3C Trace Context supports cross-service propagation. However, OpenTelemetry CI/CD conventions are release-candidate, while GenAI conventions have moved to a separate repository and many old attributes are deprecated or still development status. Pin convention versions, place unstable mappings behind an adapter, and keep Mission Control IDs as the durable domain vocabulary.

8. Personal notes and lessons learned

  • Authoritative state tells me what the factory decided; telemetry helps me understand how it behaved.
  • Observability without Mission and WorkOrder correlation becomes expensive agent wallpaper.
  • The most valuable alert identifies the human decision now required.
  • Evidence freshness is an operational SLO, not a documentation concern.

9. Interview questions

  1. How do audit events, evidence, and telemetry differ?
  2. What would you put on an operator dashboard for 20 parallel Missions?
  3. How do you trace asynchronous retries without lying about parentage?
  4. How do you observe model decisions without retaining sensitive reasoning?
  5. Which signals would cause autonomy demotion?
External review

Review this chapter.

Challenge a claim, boundary, missing failure mode, unclear term, or unsupported evidence statement.

  • Claim
  • Boundary
  • Failure
  • Evidence