Agent Engineering

Coverage

• The discipline's relationship to and distinction from prompt engineering, harness engineering, and traditional distributed systems
• The four pillars: architecture and lifecycle management, task decomposition and context management, multi-agent coordination patterns, production reliability
• The lifecycle state machine: claim → execute → verify → commit → release, with the extended research/plan/review variant for complex workflows
• Context health states (ok / degraded / compact / exhausted) and their budget thresholds, plus the six observable signals of context rot
• Multi-agent coordination patterns: orchestrator/worker, fan-out/merge, evaluator/optimiser, consensus/fusion, sequential chain, hybrid — and the cost/reliability trade-offs of each
• The two-pass pattern (audit then fresh-context implement) for reliability-critical workflows
• The eight named coordination failure modes (task stealing, context contamination, merge conflicts, silent stall, brief rot, result injection, context bloat, double-commit) with detection and mitigation
• The six production reliability requirements: observability, cost budgets, idempotency, failure recovery, safety caps, claim locks — and what breaks when each is missing
• The delegation decision framework: six gates with overhead crossover analysis (≈1000-token minimum subagent overhead), batch crossover at four tasks for cheap-model fan-out
• The most common anti-patterns (God Agent, prompt-as-architecture, memory-persisted state, runaway loop, telephone-game briefs, ghost claim) and corrective actions
• The production readiness audit checklist and the staged-rollout verification workflow (10% → 50% → 100% budget)

Philosophy

A single LLM prompt produces an answer. A system of LLMs produces a workflow that survives session boundaries, crashes, model variance, budget exhaustion, and adversarial input. Agent engineering is the discipline of building the second from the first.