AI agent

An AI agent is a system that uses a model to select actions, receives feedback from its environment, and continues toward a goal. In this encyclopedia, the term describes a pattern of control. It does not imply consciousness, unlimited independence, or reliable judgment. Definitions vary: Anthropic distinguishes model-directed agents from workflows whose paths are fixed in code. 1

The action–observation loop

A task starts with an objective and a set of available actions. The model may request an action, the surrounding software executes it, and the result becomes input for the next decision. A stop condition ends the loop. Anthropic describes environmental feedback and bounded iteration as practical ingredients of this pattern. 1

The ReAct research paper explored interleaving reasoning and actions so a language model could interact with external environments. Its experiments are evidence for particular methods on particular benchmarks, rather than proof that every agent is dependable. 2

Consider an illustrative library assistant asked to investigate a broken citation. It could open the cited page, inspect a replacement source, and propose an edit with evidence. A successful run would leave a reviewable correction and an explanation of its sources. A failed fetch should remain a failed fetch; generating a plausible replacement URL would not complete the task.

Agent or workflow?

The useful question is who chooses the next step. A workflow might always retrieve a document, summarize it, and pass the summary to a reviewer. An agent can choose another retrieval after inspecting the first result. Real systems can combine these approaches. Anthropic’s taxonomy is a helpful architectural distinction, not a universally agreed definition. 1

More autonomy also creates more possible failure paths. A fixed workflow can be easier to test when a task has predictable steps. Anthropic recommends adding complexity only when the task benefits justify the extra latency and cost. 1

Tools and authority

Tool use supplies the actions in many agent loops. With client-side tools, the model returns a structured request and the application executes it. That boundary gives the application a place to validate arguments, enforce permissions, and reject an operation. A tool request is not itself proof that an operation happened. 3

Model Context Protocol is one way to connect such capabilities. An agent can also use ordinary application functions. The protocol choice and the decision to allow an action are separate design questions.

How to evaluate a claim

For this project, we propose evaluating agents against an observable outcome: did a correction have valid evidence, did it preserve history, and did the system stop within its limits? These are editorial design criteria, not measurements from a deployed pipeline.

An evaluation should make the task, environment, allowed actions, failure cases, and budget explicit. A fluent final answer alone would not show whether a citation was checked. The long-term Agentic Bible goal includes bounded autonomous maintenance; this first slice implements reading and search only.