liveBook · Manning

AI agent

As explained in Agent Design Patterns

A software system that uses an AI model, together with prompts, data, and programmatic control, to perform tasks.

A detailed prompt can still produce an incomplete result: a model may skip a step, disregard an instruction, or drift away from the task. An AI agent addresses this design problem by placing AI models inside a software system built to perform a chosen class of tasks. The model remains central, but the system can control how work is divided, what information or tools are available, and when results are checked. section 1.0 section 1.1.1 section 1.1.2

An agent may combine several capabilities, according to its task:

  • A flexible conversational interface, often using text or voice, with repeated interaction and refinement rather than a single authoritative answer. section 1.1.1
  • Access to curated information or specialized tools, allowing it to use context or act through other software. section 1.1.2
  • A structured process that narrows the model’s responsibilities, supervises information-source choices, and supports planning or self-correction where needed. section 1.1.2 section 1.3

Turn a fragile prompt into supervised steps

Suppose a model must process many data entries but overlooks instructions and completes only part of the job. Instead of sending the entire job again, submit one entry at a time, inspect each result, and combine the accepted outputs. This does not eliminate mistakes; it makes them easier to find and contain. The same design principle motivates agent workflows that constrain model work rather than relying on prompt wording alone. section 1.0 section 1.1.2

Agents include personal assistants, coding assistants, deep-research programs, and specialized professional helpers. They commonly use large language or vision-language models, but may use smaller or specialized models. Extra orchestration is worthwhile when it adds useful information or actions, improves reliability and interpretability, or avoids unnecessarily expensive reasoning. These are design goals, not guarantees: the workflow still needs supervision when errors matter. section 1.1.1 section 1.1.2 section 1.0

How Agent Design Patterns covers it

How deep it goes
Defines agents operationally and identifies interface, model, tool, and process choices, without specifying a complete implementation.
Examples
Illustrates agents with personal assistants, coding agents, deep-research agents, specialized assistants, and entry-by-entry processing after a prompt fails.
What it assumes
Assumes familiarity with prompting a language model while defining what makes the resulting software system an agent.
What you can do after
Evaluate/understand when a structured agent workflow may improve utility, reliability, interpretability, or cost over direct model use.
How it connects
Connects AI models to agent design patterns, planning, tool usage, self-correction, and reliability.

Also explained in

Compare how 46 books explain AI agent →

Related concepts

This explainer was generated by Manning's knowledge graph from the text of Agent Design Patterns. The section links point to the parts of the book it draws on.