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/Labs/Complete source chapter
Labs2 min readlabQuick Readfailure injection lab

Orchestration Failure, Recovery, and Cost Lab

Prove bounded convergence and safe recovery under provider degradation, unknown side effects, budget pressure, duplicate delivery, partial multi step results, and human escalation. Attribute the whole cost to the accepted or failed outcome.

Status: Review readyRisk: highLifecycle: execute · verify · learnContent reviewed 2026-08-30Maturity guide →
Claim boundaryThis chapter references implementation evidence. Inspect its evidence boundary before treating a claim as proven.
Quick Read

The chapter in one pass.

~2 min

Prove bounded convergence and safe recovery under provider degradation, unknown side effects, budget pressure, duplicate delivery, partial multi-step results, and human escalation. Attribute the whole cost to the accepted or failed outcome.

Hands-on lab

Prove bounded convergence and safe recovery under provider degradation, unknown side effects, budget pressure, duplicate delivery, partial multi step results, and human escalation. Attribute the whole cost to the accepted or failed outcome.

Execute the existing Markdown instructions and retain the required output and evidence.

practice2 min chapter
Jump to validation criteria

1. Synthetic workflow

Build a three-step workflow: inspect a synthetic repository, generate a small change, and publish a mock review request. Pin workflow, model profile, tools, context, policy, evaluator, environment, attempt/tool/time/token/cost budgets, stop conditions, and approved fallback.

2. Baseline

Run the exact workflow three times. Record output quality, attempts, tool/model calls, queue and execution latency, tokens, tool/environment/evaluator cost, human review time, and accepted outcome. Establish expected error and stop behavior before injecting faults.

3. Failure injections

  1. Duplicate one dispatch command and prove lease/state/idempotency controls prevent duplicate execution.
  2. Return transient model errors long enough to open a circuit. Verify bounded jittered backoff and either approved fallback or explicit pause.
  3. Make the publish tool time out after applying a mock side effect. Verify reconciliation before retry.
  4. Feed repeated evaluator failures with no measurable improvement. Verify the loop stops and escalates instead of spending the remaining budget.
  5. Exhaust the token or monetary reservation one unit before a model call. Verify no overrun and that safe teardown/containment capacity remains.
  6. Crash the controller after a child result but before the join commit. On restart, reconcile the result exactly once.
  7. Pause and resume after changing a governing context or capability version. Verify resume is denied and a new attempt or explicit replan is required.

4. Required evidence

Retain scenario and execution manifests; admission and reservations; workflow state transitions; leases and deduplication; model/tool calls; retry, circuit, fallback, stop, pause, and reconciliation events; context/version comparison; candidate artifacts and evaluator results; human escalation and acceptance; and a ledger separating reserved, incurred, failed, recovered, human, and unallocated cost.

5. Pass criteria

ControlRequired result
Durable stateNo invalid transition, orphaned lease, or lost child result
Retry safetyNo unknown effect is retried before reconciliation
ConvergenceAcceptance or declared stop occurs inside every hard budget
FallbackOnly the pinned qualified fallback is used; changed profile is visible
ResumeGoverning drift prevents continuation under the old manifest
EvidenceReviewer can compare baseline, failures, recovery, and final outcome
CostTotal cost reconciles to reservations and exact outcome status

Any silent budget extension, unsafe retry, dropped partial result, self- accepted output, or unexplained cost is a failure.

6. Cleanup and review

Cancel remaining runs, drain queues, revoke grants, delete synthetic workspaces and mock review requests, retain the evidence bundle, and record known gaps. Repeat with the approved fallback disabled to prove safe unavailability rather than improvised routing.

Evidence boundary

Curriculum maturity is not implementation proof.

This chapter defines architecture or practice. It does not by itself prove a corresponding production implementation.

CurriculumReview readyImplementation evidenceNot asserted hereInspect evidence map →
External review

Review this chapter.

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

  • Claim
  • Boundary
  • Failure
  • Evidence