The agent systems fieldbook

Study the machinery around models.

Learn harnesses, memory, authority, recovery, and verification—then compare how real systems arrange them. Start with the complete guide or follow the question you are building around.

40 chapters · 12 labs · 1 capstone

Start with Harness Engineering Explore architectures

System trace

The work around a model

AdmitDecide whether this work may enter the system and under which authority.
ContextCompile the relevant task, user, environment, and prior state into usable context.
CapabilitiesBind the tools, permissions, and boundaries that make particular actions possible.
ExecuteRun the selected model and harness procedure against the available capabilities.
StateCommit the durable state needed to understand what was attempted and what remains open.
EffectAttempt the intended change in the external world, which may succeed, fail, or remain ambiguous.
ReconcileCompare the attempted effect with the system record and resolve any disagreement.
EvidenceCollect inspectable evidence for what occurred, rather than treating the model response as proof.
VerifyCheck whether the observed effect meets the stated conditions for success.
Accept or re-enterConsciously close the work when it is accepted, or return it to the system with the new evidence.

Start here

Three ways into the field

Read chronologically, compare real systems, or move between concepts as your own architecture takes shape.

01 · Complete guideHarness EngineeringLearn the runtime responsibilities around a model.02 · Complete guideMemory EngineeringDesign memory as a governed lifecycle, not a vector lookup.03 · Field mapArchitecture ComparisonCompare major harnesses and memory systems without a universal leaderboard.

What this guide teaches

The model is one component. The system does the work.

A useful agent needs more than a prompt and a tool call. It needs a harness that can assemble the right context, bind the right capabilities, preserve enough state to recover, handle ambiguous effects, and distinguish execution from verified completion. The guide builds that understanding one responsibility at a time.

The complete syllabus

The curriculum behind the guide

Start with the complete short guide, then use this map to go deeper. It comes from a forty-chapter working manuscript built from agent runtimes, coding agents, workflow systems, memory systems, evaluation research, and failure reports. Detailed chapters will be released as they are edited for the web.

Part I — Foundations

5 chapters

Define the harness boundary, learn the H0→H9 progression, and diagnose recurring failure classes.

Part II — Runtime mechanisms

6 chapters

Study how context, tools, environments, durable state, control structures, and memory work together.

Part III — Reliable action

5 chapters

Move from possible actions to authorized, bounded, recoverable effects in the world.

Part IV — Evidence and completion

4 chapters

Separate execution from proof, verification, acceptance, closure, and re-entry.

Part V — Comparative cases

12 chapters

Compare real systems by the responsibilities they own, exclude, and verify—not by a universal leaderboard.

Anthropic long-running harnesses Coming next
Cursor Coming next
LangGraph Coming next
Hermes Coming next
QM Coming next
Cloudflare Think and Agents Coming next
DeepSeek Harness and Cordis Coming next
Drover Coming next
Spotify Xirp and Portal Coming next
Factory Missions versus GitHub Copilot Coming next
Devin, Greptile, CodeRabbit, and Qodo Coming next
A bounded whole-product case study Coming next

Part VI — Evaluation and evolution

4 chapters

Evaluate model-and-harness configurations, inject failure, and improve the system without letting it edit its judge.

Model × harness evaluation Coming next
Conformance and trajectory evaluation Coming next
Failure injection and recovery evaluation Coming next
Harness learning, promotion, and retirement Coming next

Part VII — Course and practice

4 chapters

Turn the concepts into labs, a capstone system, review rituals, and a maintainable learning resource.

Twelve labs Coming next
Capstone Coming next
Review questions and design worksheet Coming next
Keeping the guide current Coming next

Companion guide

Memory Engineering: a practical companion guide

Design agent memory as a governed lifecycle: what a record means, where it came from, when it may influence work, how it is corrected, and when it is removed.

Read the Memory guide