Actions in other software.
Web search, code execution, email, calendar or records: the model calls tools through connectors, for example those of the MCP protocol. Some operate a computer directly, screen, keyboard and mouse, as a user would.
Understanding AI at work / Short page
Understand · Short page
Agentic AI: what are we talking about?
An agent no longer only proposes: it acts in software. This page summarises what sets it apart from an assistant, what is measured of its autonomy and the controls to plan before handing it a task.
The essentials
An agent pursues a goal by acting.
The definition, then what changes over time.
Direct answer
Agentic AI pursues a goal by choosing steps, using tools and adapting its actions to the results obtained. It can be built on an LLM.
An agent plans, acts, observes the result and starts again, for hours and, according to the providers of the most recent models, sometimes days, with intermittent human supervision.
01 / Where it sits
What sets an agent apart from other AI systems.
Four functions, four outputs, four points of attention.
| Family | What it produces | Point of attention |
|---|---|---|
| Generative AI | New content: text, image, sound or code. | Accuracy, sources, data entered. |
| Predictive AI | A score or a category. | Bias, threshold, false positives and false negatives. |
| Agentic AI | A sequence of executed actions. | Action rights, supervision, traceability. |
| Algorithmic management | An instruction or a decision. | Autonomy, appeal, intensification. |
02 / How it works
Tools, a loop, sometimes a team of agents.
The model no longer only proposes: it acts.
Act, observe, start again.
The agent plans, acts, observes the result and starts again. An error made at the start can spread through the whole chain of actions.
A coordinator and specialised agents.
One agent can direct others: it splits the goal, hands each part to a specialised agent, then assembles the results. Agents built on the same model share the same blind spots.
Provider figures · November 2026
Anthropic states that Claude Fable 5.2 can coordinate hundreds of sub-agents, Google offers a “swarm” mode with Gemini 4 Carbon, and OpenAI says it used about 10,000 agents for 88 hours for a mathematical proof that has not been independently verified. These figures say what is possible, not what is reliable.
03 / Measurements
What is measured of their autonomy.
Two dated reference points, with their limits.
| Reference | What it measures | Result | Limit |
|---|---|---|---|
| Time horizon · METR | The length, in working time of a human expert, of the tasks an agent succeeds at half the time. | From 3 seconds for GPT-2 in 2019 to 17 h 25 for Claude Mythos Preview in April 2026; a doubling roughly every seven months from 2019 to 2025. | Mostly software tasks; measurements unreliable beyond 16 hours; series stopped on 8 September 2026. |
| Remote Labor Index · CAIS and Scale AI | The share of 240 real freelance projects whose deliverable is as good as the professional’s. | 2.5% at launch in October 2025, 41% with Claude Fable 5.2 in November 2026. | Nearly six projects in ten remain out of reach; the test excludes the client relationship, physical work and long-running projects. |
METR, Time Horizon 1.1 ↗ · Remote Labor Index ↗ · Guide, chapter 07
Worth remembering
A completed task is not an automated job. A job combines tasks, relationships, responsibilities and the unexpected.
04 / Controls
Four risks, four controls.
The more agents there are, the more control must bear on rights, the log and the stopping points.
| Risk | Control to plan |
|---|---|
| An initial error spreads through the whole chain of actions. | Stopping points, a log of actions and a way to roll back. |
| A document read by the agent contains hidden instructions: this is prompt injection. | Limit the rights granted, action by action: the risk rises as soon as the tool can send, modify or delete. |
| Agents built on the same model reproduce the same error. | Make control bear on the rights granted, the log of actions and the stopping points: dividing up the work is not checking it. |
| Human control shrinks to a rubber stamp. | Check that the person who validates has the time, the information, the skills and the authority to refuse. |
OWASP, LLM01:2025 ↗ · Guide, chapter 08
An incident · July 2026
According to published accounts, several hundred OpenAI agents taking part in a cybersecurity exercise broke into Hugging Face systems. Read the analysis of this incident
05 / At work
Supervising an agent is a task.
Human control has working conditions too.
Supervising an agent is not an abstract safeguard: it is a task given to an employee, with a workload, constraints and a responsibility. When validation requests follow one another, users end up no longer reading what they approve: this is approval fatigue.
The scale in the works council dossier places a system from A0 (information) to A4 (delegation), through recommendation, supervised action and autonomous action. The higher the level, the more human control needs proper tooling.
Human in the loop: what supervising does to the supervisor · Works councils and AI: qualifying autonomy
Key answers
Three frequent questions.
Short answers, drawn from the guide.
What is the difference between an AI agent and a chat assistant?
The assistant produces content that you review. The agent itself chains actions in software to reach a goal: it no longer only proposes, it acts. Attention shifts from the accuracy of the text to action rights, supervision and traceability.
Can an AI agent act without human validation?
It depends on the rights it is given. At levels A3 (autonomous action) and A4 (delegation) of the works council scale, it acts without prior validation. You then need to define in advance what it may do alone, require explicit validation for actions with serious consequences, keep a log of actions and keep a real way to stop it.
What is a multi-agent system?
An architecture in which a coordinating agent splits the goal, hands each part to a specialised agent, then assembles the results. The main providers now offer it. Because the agents often rely on the same model, the same error can repeat from one to the next.
Where this page comes from
It summarises one chapter of the guide Understanding AI at work, which gives the source, date and limits of every figure. The figures dated November 2026 will go out of date: the guide is the version kept current.
Update log
- First version, drawn from the guide “Understanding AI at work” updated on 13 November 2026.