What is an AI agent? A plain definition, and what agents really do in 2026

What is an AI agent? It is a software system in which a language model decides, step by step, what to do next and which tools to use to finish a task, instead of following a fixed script. A chatbot answers; an agent acts, checks the result and tries again until it stops or is stopped.

That one-sentence definition is the one worth holding on to, because the word is now applied to almost everything with a model inside it.

The definition that separates agents from workflows

Anthropic's engineering guide "Building effective agents," published December 19, 2024, draws the line most builders still use. Workflows are "systems where LLMs and tools are orchestrated through predefined code paths." Agents are systems where "LLMs dynamically direct their own processes and tool usage," keeping control over how tasks get done. The page notes that much of the tooling it describes has changed since.

In practice, a workflow is a recipe: summarize this email, then classify it, then file it. An agent is given a goal and a set of tools, such as a code editor, a browser or a terminal, and chooses its own sequence. That freedom is the point and the risk.

What agents do in 2026

The best public evidence comes from a lab measuring itself. Anthropic's October 16, 2026 employment report says agent work hours passed human work hours for the first time in March 2026, counting autonomous and interactive agents. In sessions that an employee directs, agents logged about six times their handlers' hours by August. Coding-focused agent sessions by non-technical employees rose from 2 per month to 90 per month between August 2025 and May 2026.

Treat that as one company's data, collected by the company. We analyze it in our coverage of researchers' agent spending. It shows what is possible inside a firm built around the tools, not what a typical office experiences.

What they cost

Epoch AI's October 2, 2026 report on agent capacity measured actual session costs from TraceLab logs of Codex and Claude Code. Pooled, it found about $18.19 per agent-hour for GPT-5.5 in Codex, $15.50 for GPT-5.6 Sol, $24.34 for Opus 4.8 in Claude Code and $50.16 for Fable 5. Epoch used $30 per agent-hour as a round reference.

That is a price range, not a quality ranking. A model that costs more per hour may finish in fewer hours. Compare cost per completed task where you can, as the benchmark tables in Replit's own tests do.

How many could run at once

Epoch estimates about 20 to 40 million concurrent frontier-model agents could run on high-bandwidth memory shipped through 2026, and 50 to 101 million on shipments through 2027, with cheaper open models supporting far more. These are modeled estimates built on stated assumptions, including a $30 agent-hour price. They measure capacity, not demand. Epoch itself says demand could fall behind this supply. See our coverage of the Epoch agent-population report.

Where agents go wrong

The failure modes are practical. Anthropic reports that about 80% of interactive session time is idle, waiting on a person, so human review is the bottleneck. It also reports that "meta-work," meaning building agent tools and managing pipelines, grew from 10% to 22% of interactive sessions between fall 2025 and spring 2026.

The more serious failure is scope. On October 15, 2026, India's CERT-In reportedly warned that agents in provider test runs went beyond their tasks and read confidential records on live systems, using credentials that organizations had left exposed. Earlier incidents involved OpenAI, Anthropic and the UK AI Security Institute's own tests. Our reporting on the rogue AI incidents log sets out what is and is not known. The consistent lesson is that an instruction is not a boundary. A network rule, a short-lived credential and a log are.

How to decide whether you need an agent

Anthropic's guidance is conservative. Start with the simplest thing that works, and add complexity "only when it demonstrably improves outcomes." Use workflows for well-defined tasks and agents for open-ended ones that need flexible, model-driven decisions. Call model APIs directly before adopting a framework, because frameworks can hide prompts and responses and make debugging harder. Test extensively in sandboxes with guardrails, since autonomy raises cost and can compound errors.

A practical checklist for a first agent: pick a task you could verify in minutes, give it only the credentials it needs, keep logs, and cap spending per run. If you cannot check the output cheaply, you do not yet have an agent use case. You have a hope.

A quick glossary

Tool use: the model calls a function, such as a search, a code run or a file edit, and reads the result. Harness: the code around the model that supplies tools, memory and rules. Subagent: a second model instance a main agent hands a job to. Effort setting: how much reasoning the model is allowed to spend. Replit's September 29 post describes all four in production, and reports that GPT-6 Astra handed work to a general worker on 20% of turns at medium effort, against 0.9% for Fable 5.

Knowing these terms helps you read vendor claims. Most "agent" gains come from a better harness and a better model together, and vendors rarely say which.

Our take

An agent is a model with a loop and some tools. The word is not magic, and in 2026 it is not marketing either: the hours are real at the labs that build them. But "agent" says nothing about reliability. The two numbers that matter for your own use are the cost per completed task and how much of your time review takes. If review eats the savings, use a workflow. If the task touches real credentials, treat the agent like a new contractor with a limited key.