Agents · Intermediate
AI Agents
Also known as: LLM agents, Agentic AI
In an LLM agent, a model chooses actions or tools, observes their results and decides what to do next within a defined task.
What AI Agents is
Agent has several meanings. This page concerns language-model systems. Anthropic distinguishes a workflow with predefined code paths from an agent that dynamically directs its process and tool use. A multi-step workflow is not necessarily an autonomous agent.
How it works
ReAct studies interleaving reasoning traces with task-specific actions. Actions can retrieve information from an external environment; the result informs the next step. This is a research pattern, not a guarantee that a generated plan or action is correct.
Why it matters
The useful design question is which decisions need model judgment and which should remain explicit rules. A fixed process may be sufficient when the required steps are already known.
Common uses
- →Compare a fixed workflow with a model-directed sequence
- →Design a bounded, read-only research task
- →Identify where a proposed tool action needs a person's approval
Strengths
- ✓Can choose subsequent steps using tool feedback; success depends on the task and implementation
Watch for
- ✓A plausible plan can still be wrong
- ✓Additional calls can add latency and cost
- ✓A tool response is evidence to inspect, not permission to take further actions
Proposed read-only invoice exercise
Fictional invoice INV-07 says total 120 USD and has no due date. The permitted tool can read that invoice; it cannot send messages, pay or edit records.
- Ask for the invoice ID, total, currency and due date, with a source location for each field.
- Allow the model to request the document only. Keep the expected answer outside its input.
- Check for INV-07, 120 USD and due date unknown. Stop if it requests an unapproved tool or invents a date.
- Record failures, number of calls and human corrections before deciding whether a fixed extraction workflow is enough.
A proposed result to check, not captured model output: INV-07; 120 USD; due date unknown. This example has not been run by Moyan.
Continue exploring
More in this collection
Browse all AI ConceptsSources & References
Anthropic — Building effective agents (2024): workflows, agents and evaluator–optimizerYao et al. — ReAct (2023 revision): reasoning, actions and observations