Agent Engineering

Core Philosophy

An agent is an LLM that decides which tools to call, in what order, with what arguments, based on intermediate results — looping until the task is complete. Every loop iteration costs tokens, latency, and money. Every unnecessary tool call, every bloated prompt, every redundant context injection is waste that compounds across thousands of executions.

The engineering discipline is: accomplish the task in the minimum number of LLM calls, with the minimum tokens per call, using the minimum tool invocations, while maintaining reliability. These four objectives are not in tension — wasteful agents are also unreliable agents, because every unnecessary step is another opportunity for the model to hallucinate, lose track of its goal, or choose the wrong tool.

The Agent Performance Triangle

Every agent design decision involves three competing forces:

Reliability — does the agent accomplish the task correctly? This is the constraint, not the optimization target. Establish a reliability floor first (e.g., 95% correct on representative tasks), then optimize cost and speed without dropping below it.

Cost — how many tokens does the agent consume? Token cost = input tokens + output tokens across all LLM calls. Cost scales linearly with loop iterations and context size.

Latency — how long does the task take end-to-end? LLM inference time scales with input token count (time to first token) and output token count (generation time). Sequential tool calls add latency linearly; parallel tool calls add latency once.

The highest-leverage optimizations improve all three simultaneously: fewer loop iterations means lower cost, lower latency, AND fewer chances to go off-track.