Agency is the ability to perceive a situation, decide what to do, and act with some autonomy toward a goal. A prompt-only assistant can produce a response and stop. An agent can instead work through multiple steps, using each result to choose its next action. The distinction is what the system can decide and do over time, not whether it ever asks permission. section 1.1 section 3.2.4
A useful mental model is the sense–plan–act–learn cycle:
- Sense: receive a goal, context, or feedback. section 1.1.2
- Plan: decide which tasks and tools the goal requires. section 1.1.2
- Act: execute a tool for the current task. section 1.1.2
- Learn: inspect the result, then decide whether to stop or continue. A tool’s output may feed the next call. section 1.1.2
Table 1.2 Example goals, and the tasks and tools needed to carry them out
| Goals |
Tasks |
Tools |
| Create an image |
Create an image |
|
| Travel to Calgary (YYC) |
Search for flights Book flights Book hotels Book transportation |
|
| Buy a computer |
Search for computers Compare features and prices Order a computer |
|
Compare an image request with travel booking or computer buying: a larger goal calls for several tasks and different tools, rather than one response. section 1.1.2
From prompt step to agent
A research planner that only generates a plan is a prompt step, even if its output is useful. Giving it a curated set of tools lets it decide when and where to act. Tracing tool calls can then reveal its decision path; typed output schemas help keep results inspectable. section 2.4 section 2.6
More agency also creates more ways for an error to propagate: one poor tool choice can affect later steps. Start with the least autonomy the goal needs. Limit available tools, validate outputs, cap action budgets, and require a human checkpoint for irreversible actions. Loosen those boundaries only as the agent demonstrates reliable behavior; higher-stakes work warrants tighter limits. section 3.2.4 section 4.1.3
