Agent Engineering¶
A dedicated track for everything involved in shipping autonomous AI systems — not scattered across modules, but organized as one engineering discipline.
flowchart TB
subgraph Loop["1 · Agent Loop"]
P[Perceive] --> R[Reason]
R --> A[Act]
A --> O[Observe]
O --> P
end
subgraph Runtime["2 · Harness"]
Perm[Permissions]
Term[Termination]
State[State]
end
subgraph Cap["3 · Capabilities"]
Mem[Memory]
Tools[Tools / MCP]
RAG[RAG]
end
subgraph Scale["4 · Orchestration"]
Sup[Supervisor]
Work[Workers]
Hand[Handoffs]
end
subgraph Quality["5 · Quality"]
Trace[Tracing]
Eval[Evals]
Mon[Monitoring]
end
Loop --> Runtime
Runtime --> Cap
Cap --> Scale
Runtime --> Quality
Read in order¶
| # | Topic | Page | Also in courses |
|---|---|---|---|
| 1 | Agent loop | The Agent Loop | Course 07 L1–3 |
| 2 | Memory | Memory Systems | Course 07 L5 |
| 3 | Tools & MCP | Tools & MCP | Course 08 L3–4 |
| 4 | Harness engineering | Harness Engineering | Course 08 |
| 5 | Orchestration | Orchestration | Course 09 |
| 6 | Observability & tracing | Observability & Tracing | Course 08 L6, Course 12 |
| 7 | Agent evals | Agent Evals | Course 13 L4 |
What is harness engineering?¶
Harness engineering is the discipline of building the runtime around an LLM agent — not the model, not the prompt alone, but the system that makes agents reliable:
| Primitive | What it does |
|---|---|
| Loop | Perceive → reason → act → observe until done |
| State | Checkpoint conversation, tool results, plan |
| Tools | Sandboxed execution, schemas, MCP servers |
| Permissions | Allowlists, human-in-the-loop, budget caps |
| Termination | Max steps, success criteria, timeout |
| Observability | Spans per step, token/cost attribution |
| Evals | Trajectory regression, tool-call correctness |
Inspired by Awesome Harness Engineering and Agents Towards Production.
Quick reference¶
| I need to… | Go to |
|---|---|
| Understand ReAct | Agent Loop |
| Persist context across sessions | Memory |
| Connect to APIs / filesystem | Tools & MCP |
| Make a coding agent safe | Harness Engineering |
| Run multiple specialists | Orchestration |
| Debug a failed run | Observability |
| Gate a release | Agent Evals |
Related¶
- 2026 Skills — Claude Code, skills, loop engineering
- Evals & Observability hub
- Glossary — harness, MCP, trajectory eval
- Related papers — ReAct, Toolformer, MemGPT, AgentBench, and more