Living book · v0.76.0 · DOI 10.5281/zenodo.21632412
Harness Engineering
A living book on the scaffolding that surrounds AI agents
🗞 Living-book Radar · 2026-08-07 · impact B · item in Portuguese
⭐⭐ Meta lança o Muse Code (05/ago) — agente de terminal com subagentes de fundo persistentes, cada um em workspace isolado, e event log local à prova de crash, movido pelo Muse Spark 1.2. Fechado: binário proprietário por script, sem repositório público. Cobertura de 1ª linha (Bloomberg · CNBC · The Register); ⏳ primária da Meta bloqueada pelo egresso
see the full Radar →📖 This edition (v0.76.0 · 2026-08-07): o endereço do site vira variável — History
Continue reading Resume ▶
Track · 1FoundationsThe book's vocabulary and thesis.
Track · 2CapabilitiesThe 16 components, one per chapter.
Track · 3BenchmarkReal harnesses, compared.
Track · 4Hands-onBuild harness-zero, step by step.
Opening
00IntroductionWhat a harness is and why it matters.01FoundationsVocabulary, the expiration clause, a map of the territory.
Chapters by capability
02The Agent LoopThe cycle that turns a chat into an agent.03Context DeliveryAssembling the right context at the right time.04CompactionFitting the window without losing the thread.05Tool DesignThe contract between the model and the world.06MCPThe protocol that plugs in external tools.07Permissions & SandboxingPower with containment: the golden rule.08Memory & StateWhat the agent remembers across turns and sessions.09PlanningThinking before acting, when the task calls for it.10Subagents & OrchestrationDivide and conquer for large tasks.11Verification & EvalsHow to know the agent got it right.12ExtensibilityHooks, plugins, and a harness that grows.13InterfacesCLI, IDE, web: where humans meet the agent.14Convergences & TrendsWhere the discipline is heading.15The Embedded HarnessWhen the harness becomes the product.16Learning & Self-ImprovementA harness that learns from use.17The Protocol LayerThe standards that bind the ecosystem.
Benchmark · Apparatus · About