Agent optimization is the work of increasing decision quality under real constraints: incomplete evidence, tool latency, cost, and a user who needs to know what happened.
Figure 02 — the optimization loop
flowchart LR G[User goal] --> X[Context packet] X --> D[Agent decision] D --> A[Action] A --> R[Observed result] R --> J[Judgment] J --> X J --> D
A good target is observable
“More helpful” is not a target. A useful target describes the decision, the evidence required, the action boundary, and the condition for success. It also preserves negative cases—moments when the right behavior is to ask, defer, or stop.
Begin with a small set of representative tasks. For each, record the intended result and why it is correct. This makes evaluation a design tool rather than a retrospective argument.
Improve one lever at a time
- Improve the task framing when the agent pursues the wrong objective.
- Improve the available evidence when answers sound plausible but miss facts.
- Improve tools when the agent knows what to do but cannot complete it.
- Improve the evaluator when the system cannot distinguish useful from merely fluent.
Prompts are an interface, not a moat. Durable gains usually come from a stronger context packet and a harness that can replay the work.
Read next: Harness engineering: make every run inspectable →
For annotated video and curation, enter For Collaborators →