Learn how modern AI agents are built around the LLM.
English · 繁體中文 · 简体中文 · 日本語 · 한국어
The model reasons. The harness turns that reasoning into controlled action: it runs tools, keeps state across calls, gates side effects, and coordinates loops. A model call cannot do any of those things by itself.
This repo explains the harness section by section: loop, tools, memory, permissions, context, tasks, and interfaces. Learn it once and you can read many agents, since a coding tool, chat assistant, and autonomous runner mostly differ in harness choices.
Three companion repos go deeper than one section can:
- learn-agent-memory: scales the memory loop into a production memory subsystem.
- learn-deepseek-harness: learns deepseek-harness from scratch, one plugin seam at a time.
- EvalGrill: turns real cases from agent applications into reproducible, verifiable eval sets.
Contents: Loop · Method · Systems · Sections · Structure · Running
Most agents share the same control flow: call the model, run requested tools, append results, and call the model again.
The loop is small. Most engineering is around it: dispatch tools, gate side effects, manage context, persist state, and coordinate other loops.
Every section is self-contained and uses the same four-part lens:
- Opening. What problem this layer solves.
- Mechanism. The general design and control flow.
- Per system. How real systems implement it.
- Failure modes. What breaks and how to mitigate it.
To learn from this repo:
- Read the sections in order. Each builds on the layer before it.
- For a runnable section, read
src/loop.py, then run itsdemo.py. - Diff a section's
src/against the section before it. The diff is the one mechanism that section adds.
Each system is a worked example for the sections below.
| System | Why people use it | Read it for | Sections | Version studied |
|---|---|---|---|---|
| Claude Code | Frontier coding agent: edits files, runs commands, ships changes in real repos. | The full harness, start here | 0 to 23 (all) | v2.1.88 |
| Hermes Agent | Long-term assistant: remembers you, learns workflows, runs anywhere. | Memory, skills, always-on channels | 7, 9, 14, 16, 19, 21, 22 | v2026.7.1 |
| mini-swe-agent | Research baseline: one bash tool, about 150 lines. | The smallest complete loop, budgets, eval harness | 0 to 3, 8, 10, 11, 20 to 23 | v2.4.5 |
| deepseek-harness | Plugin-first harness: even the loop is a replaceable plugin. | Plugin seams, durable session log, ACP | 1 to 8, 10 to 14, 16 to 21 | dsh-v0.1.0-rc.7 |
| (more soon) |
More systems can be added later, including OpenClaw and aider. Two companion repos go deeper: learn-agent-memory for the memory layer, and learn-deepseek-harness for learning deepseek-harness from scratch.
Eight layers, from the basic loop to a harness that runs itself. Each row links to one self-contained writeup.
Section 9 continues in learn-agent-memory: ten more stages that scale its memory loop to production.
| # | Section | Question | Key mechanisms |
|---|---|---|---|
| Layer 0 · Foundations | |||
| 0 | Harness thesis | Where does agency come from? | Model vs harness, actions, observations, permissions |
| Layer 1 · Core Loop | |||
| 1 | Agent loop | How does an agent keep going? | messages[], loop, stop_reason |
| 2 | Tool runtime | How are tools called and routed? | Registry, schemas, dispatch, deferred search |
| 3 | Permission & sandbox | How are side effects gated? | Permission modes, approvals, sandboxing |
| 4 | Hooks | How do extensions attach to the loop? | PreToolUse, PostToolUse, lifecycle events |
| Layer 2 · Complex Work | |||
| 5 | Planning & todos | How is big work decomposed? | Plan mode, todo list, approval before edits |
| 6 | Subagents | How is a subproblem isolated? | Fresh messages[], delegation, child loop |
| 7 | Skills | How are capabilities loaded on demand? | SKILL.md, catalog, progressive disclosure |
| 8 | Context management | How do long sessions fit the window? | Budgeting, stubs, compaction, summaries |
| Layer 3 · Knowledge & Resilience | |||
| 9 | Memory | How does it remember across runs? | Selection, recall, extraction, consolidation |
| 10 | System prompt assembly | How is the prompt built each turn? | Prompt sections, live state, cache boundaries |
| 11 | Error recovery | How does a long task survive failure? | Retries, overflow recovery, fallback model |
| Layer 4 · Long Running & Async | |||
| 12 | Task system | How does work persist beyond a turn? | Task records, dependencies, locks |
| 13 | Background execution | How does work run off the main loop? | Handles, task state, notification queue |
| 14 | Scheduling | How does an agent run later? | Cron, sleep, remote triggers, queues |
| 15 | Worktree isolation | How does parallel work avoid collisions? | Git worktrees, cwd binding, safe cleanup |
| Layer 5 · Multi Agent | |||
| 16 | Coordination | How do many agents talk? | Inboxes, broadcasts, permission bubbling |
| 17 | Protocols | How do agents agree and stop cleanly? | Plan approval, shutdown handshakes |
| 18 | Autonomy | How do agents organize themselves? | Idle cycle, task claiming, self organization |
| Layer 6 · Extension & Integration | |||
| 19 | MCP / plugins / channels | How does the harness reach the world? | Transports, channels, tool pool assembly |
| 20 | Observability & evaluation | How do we know it works? | Tracing, metrics, evals, failure analysis |
| 23 | Evaluation | How do we know a change made it better? | Eval environments, resets, judges, Pass^k |
| Layer 7 · Composition | |||
| 21 | Loop engineering | How do loops stack into a system that runs itself? | Verification loop, triggers, budgets, maturity levels |
| 22 | Graph engineering | When does control flow move from the model to code? | Nodes, coded edges, cycles, agent nodes, typed decision edges |
All 24 section writeups are present, from 00-harness-thesis/ through 23-evaluation/.
learn-agent-architecture/
├── README.md # top-level map
├── sections/ # one folder per section
│ ├── 00-harness-thesis/ # README.md per section
│ ├── 01-agent-loop/src/ # runnable chain starts here
│ ├── ...
│ └── 23-evaluation/
└── references/ # primary sources and prior art
Each section folder is NN-name/ and contains a README.md.
Sections 1 to 23 also carry a runnable src/. The code accumulates section by section.
Each section adds one mechanism and evolves loop.py, so a diff between adjacent sections shows what changed.
Deep dives that outgrow one section live in their own repos. learn-agent-memory scales the section 9 loop into a full memory subsystem. learn-deepseek-harness learns deepseek-harness from scratch, one plugin seam at a time.
Sections 1 to 23 ship runnable demos. Set up once from the repo root:
uv venv
uv pip install -r requirements.txt
cp .env.example .env # then add your ANTHROPIC_API_KEYPinned dependencies are in requirements.txt. .env is gitignored and holds:
ANTHROPIC_API_KEY- optional
ANTHROPIC_MODEL - optional
ANTHROPIC_BASE_URL
Each runnable section has:
test.py: offline checks, no key needed.demo.py: live demo against the API.
python sections/01-agent-loop/src/test.py # offline
uv run python sections/01-agent-loop/src/demo.py # live- Add a system. Slot a new agent into the same section structure.
- Deepen a section. Add a mechanism, clearer diagram, or sharper failure mode.
- Correct the record. These are reconstructions from source, docs, and behavior. Sourced corrections are welcome.
Favor named, verifiable mechanisms over speculation. Cite sources. See CONTRIBUTING.md for the full PR checklist.
- claude-code: Claude Code source backup used for mechanism names and implementation paths.
- hermes-agent: Open-source agent harness (MIT) used as the second system under study.
- mini-swe-agent: Minimal SWE agent (MIT) used as the third system under study.
- deepseek-harness: Plugin-based agent harness (MIT) used as the fourth system under study.
- learn-claude-code: Code-first harness reconstruction and section framing.
- EvalGrill: companion tool (Apache-2.0) that applies section 23, building eval sets from real agent-application cases.
- Anthropic Agent Skills best practices: Progressive disclosure levels for skills.
- Anthropic prompt caching: Cache breakpoints, TTLs, pricing, and token minimums.
- cobusgreyling/loop-engineering: Loop building blocks and readiness levels.
- LangChain · The art of loop engineering: The four stacked loops.
- Addy Osmani · Loop engineering: Composed building blocks for agent loops.
- MindStudio · What is loop engineering: Goal conditions for autonomous workflows.
- Lilian Weng · Harness engineering for self-improvement: The improvement loop, with gates outside the loop.
- LangChain · 3 years of graph engineering: Nodes, edges, cycles, and agents as nodes.
- Anthropic · Building effective agents: Workflows vs agents and the five workflow shapes.
- Google · Why we built ADK 2.0: Routing in code and context isolation between nodes.
- TypeSafe Jev docs: the request contract and answer shape of a typed decision model.
Section 22 wires one into a graph edge in
decide.py. - ai-agent-book: 《深入理解 AI Agent》 by 李博杰 (Apache-2.0). Chapter 6 grounds the evaluation section.
Thanks to these collections for listing this project: