What an Agent Is

A language model produces responses. An agent takes actions: it reads files, runs commands, calls services, and decides what to do next until the work is complete or a stopping condition is reached.

Agent = Model + Harness

The model supplies language and reasoning capability. The harness supplies the operating environment: tools, context, permissions, loop control, and failure handling. Research supports treating these as independent contributions to success. Better scaffolding can improve results without changing the model, while a weak harness wastes a strong model.

When an agent misbehaves, inspect the complete system before blaming the model. Part II develops the harness in detail, while the Agent-Computer Interface focuses on the boundary between reasoning and action.

Five design levers

Everything you can change about an agent falls into five places:

1. Instructions: how the task, constraints, and success conditions are stated. Covered next in Structuring Agent Instructions.
2. Model: the capability ceiling and the cost floor. Covered in Working With Probabilistic Models and Choosing a Model.
3. Context: what the agent knows at the moment it acts. Introduced here, treated fully in Context.
4. Tools: what actions are available and how they are described. Covered in Designing Agent Tools.
5. Harness: the execution environment that contains the other four. The subject of Part II.

Model choice sets a capability ceiling, but it does not determine the result alone. Two systems using the same model can behave very differently because they supply different context, tools, constraints, and feedback. Agentic engineering works on all five levers as one system.

The agent is the whole system

This definition prevents a common diagnostic mistake. If an agent misses a requirement, the cause may be a weak model. It may also be missing context, an ambiguous tool, an instruction conflict, or a loop that stopped before verification. Changing the model without locating the failed lever makes the system more expensive without making it more reliable.

The next chapter begins constructing that system with its most immediate control surface: the agent’s instructions.