Five layers, five different jobs
Select a layer. Each is a distinct lever; improving one cannot fully compensate for a broken neighbor.
Watch one request move through the system
Choose a scenario, then step through the five layers to see how responsibility is divided.
Know which lever you are pulling
The same symptom can tempt the wrong fix. This table makes the boundaries explicit.
| Layer | Changes | Best for | Not enough when… |
|---|---|---|---|
| Model | Base reasoning, language, perception | Broad capability ceiling | The agent lacks your facts or access |
| Context | What the model knows now | Relevance and personalization | The agent must take an action |
| Skills | Repeatable procedures and tool use | Domain reliability | Steps need orchestration or recovery |
| Harness | Control flow, state, permissions, retries | Multi-step execution | The underlying judgment is weak |
| Evals | Measurement and feedback loops | Knowing what improved | No one acts on the evidence |
Where should you invest next?
Pick the failure pattern closest to what you observe. The tool points to the likely first intervention—not necessarily the only one.
A practical build order
Start with the narrowest complete loop. Add autonomy only after the loop is observable and safe.
- 1 · DefineWrite the job as an observable outcome: inputs, output, boundaries, and what “done” means.
- 2 · GroundProvide the minimum context needed for this run. Separate stable policy from task-specific facts.
- 3 · EquipAdd one skill at a time. Give every tool a clear contract: when to call it, inputs, outputs, and failure modes.
- 4 · OrchestrateBuild the harness around the happy path, then add permission gates, timeouts, retries, and human review where consequences rise.
- 5 · EvaluateTest a small, representative set—including edge cases. Measure task success, not how impressive the prose sounds.
- 6 · ImproveFix the cheapest responsible layer first. Upgrade the model only when the ceiling—not the scaffolding—is the bottleneck.
Can you locate the failure?
Four quick cases. Choose the first layer you would improve.
The shortest useful summary
Use this distinction in design reviews and debugging sessions.
It sets the baseline capability and reasoning ceiling.
It supplies the right facts, instructions, and current state.
They package repeatable domain procedures and tools.
It runs the loop, manages state, and contains risk.
They show where the system fails and whether a change helped.
They set intent, approve consequences, and own the outcome.