Harness
Everything around the card: reader, token vault, blocklist, transaction logs.
- Few tools, but well defined
- Least-privilege permissions
- Saved state, easy to resume
- Easy to pause and audit
Choose your lens
A good AI model is not enough
For an agent to work in the real world, three layers must be designed well: environment, feedback, flow. Think of open payment on a transit network. The bank card is the model. The rest is everything around it.
Agent
Agent = Model + Harness ( Graph ( Loop ) )
Model
The bank card: the capability
HarnessThe validator and back-office: reader, token vault, blocklist, logs
GraphThe fare journey: tap-in, transfer, tap-out, daily cap
LoopTap, authorize, open or refuse. Inside one or more steps.
First, what is the model?
The model is the AI itself: a system trained on huge amounts of text that reads what you give it and writes a response. It is the brain of the agent, not the whole agent.
Agent = model + harness
What it does well
- Understands instructions in plain language
- Writes, summarizes, reasons, plans
- Proposes the next action to take
What it cannot do alone
- Remember a project from one session to the next
- Run a test or open a browser
- Enforce permissions or approvals
- Restart a job that failed
In the analogy
The model is the bank card: it carries a real capability (it can pay), but it cannot open a gate, compute a fare or handle a refusal on its own. It needs the validator and back-office around it, the authorization cycle to decide, and the fare journey to tie it all together. The three layers turn a card into a working transit payment.
The three layers
Each one answers a different engineering question. They are not synonyms.
Loop
Tap, authorize, decide, retry within limits.
- Loop on evidence, not on confidence
- Clear, measurable goal
- Capped number of attempts
- Plan B: escalate to a human
Graph
Linked steps: tap-in, transfer, tap-out, daily cap, bank settlement.
- Useful with several specialists
- Useful for approvals and recovery
- Overkill for simple tasks
- Brittle if the plan keeps changing
Not working? Fix the layer that owns the failure
Diagnose first. Blaming the model is rarely the answer.
| The symptom | Start with |
|---|---|
| No safe access to the right tool or data | Harness |
| Forgets progress between sessions | Harness |
| First attempt is close but unreliable | Loopadd an external check |
| Keeps going after success, or stops without proof | Loopevidence-based stop rule |
| Several specialists must run in a set order | Graph |
| Workflow changes too often for a fixed diagram | Keep it simpledelay the graph |
The expensive mistakes
Five ways teams waste money on agent architecture.
Graph before understandingDrawing 40 nodes before observing how the agent actually solves the task.
Self-gradingLetting the same model write and grade its work, with no independent check.
"Keep trying" loopsLooping until it works, with no goal, no cap and no escalation path.
Harness as a dumping groundPiling on tools and memory. More options means more wrong choices.
Blaming the modelStale state, vague tool schemas and missing exit conditions are orchestration failures, not intelligence failures.
What to remember
The three layers do not replace each other. A beautiful graph with weak memory is useless. A great harness without a stop rule burns money. Good loops become unmanageable when branching and approvals are hacked together. Reliable agents come from designing all three together.