Agent Harness Tutorial · Architecture
A north star, not a checklist
Everything in this tutorial together is a north star: the whole corpus of what a harness can need. Most agents need only a slice of it. When a knife is enough, don't forge a sword.
Eight precautions: start from the archetype; scale by risk; deterministic first; reuse, don't rebuild; name the delta; leave out on purpose; add on evidence; review for excess too.
Ten shapes of agent. The shape, not the domain, decides what an agent owes. Two questions sort them: who controls the flow, and what does the output touch?
| ID | Shape | Who controls the flow | Example |
|---|---|---|---|
| A1 | Single-Turn Transform | our code | summarise a claim note |
| A2 | Structured Extractor | our code | claim form into fields |
| A3 | Grounded Answerer | our code | policy wordings |
| A4 | Conversational Assistant | our code, across turns | broker desk Q&A |
| A5 | Deterministic Workflow | our code (the graph) | renewal reminder pipeline |
| A6 | Tool-Using Agent | the model | claims FNOL intake · policy servicing |
| A7 | Multi-Agent System | several models | FNOL → fraud signals |
| A8 | Artifact Producer | the model | SQL for an actuarial report |
| A9 | Autonomous Operator | the model, nobody watching | nightly chase of missing documents |
| A10 | Classifier · Router · Judge | the model's verdict | claims triage router · fraud flag · LLM judge |
Shapes are tags, not a tree. Claims FNOL is A6 and A4; with a hand-off to fraud signals it would also be A7. It owes the union.
Want every detail? The long version is in the Series: a word of caution: the north star.