agent-engineering-expert

Agent Engineering Expert

An agent is a loop in which a model chooses tools until a goal is met. The difficulty is not the loop — it is that the model's choices are driven by text that may be attacker-controlled, and the tools have real effects.

Core Concepts

The Loop

observe → decide → act → observe …

Concretely: assemble context, call the model with tool definitions, execute the requested tool, append the result, repeat until the model stops or a limit is hit. Every element of that loop needs a bound — iterations, wall time, tokens and cost.

Capability Is the Design

What an agent can do is defined entirely by the tools it holds. Reliability and safety come from designing that surface, not from instructing the model to be careful. A tool the agent does not have is a failure mode that cannot occur.

The Model Is Untrusted Input

Anything the agent reads — a web page, a document, a tool result, an email — may contain instructions aimed at the model. Treat every tool call as a request from an untrusted party and authorise it independently. This is the defining security property of agents, and no amount of prompting removes it.

Autonomy Is a Dial

More steps means more compounding error. A workflow with fixed stages and a model at each decision point is usually more reliable, cheaper and easier to debug than an open-ended agent. Choose the least autonomy that solves the problem.

Designing Tools

The tool description is a prompt. It is the only thing telling the model when to use it.

{
  "name": "search_orders",
  "description": (
      "Search a customer's orders by status and date range. "
      "Use when the user asks about their order history or the state of an order. "
      "Returns at most 50 orders, newest first. "
      "Does not return payment card details. "
      "Use `get_order` instead when you already have an order id."
  ),
  "input_schema": {
      "type": "object",
      "properties": {
          "status": {"type": "string", "enum": ["pending", "shipped", "cancelled"]},
          "since":  {"type": "string", "format": "date"},
          "limit":  {"type": "integer", "minimum": 1, "maximum": 50, "default": 20},
      },
      "required": ["status"],
  },
}