Concept

Agentic AI

AI that can take sequences of actions on its own to complete a goal - planning, using tools, checking its own work, and iterating without needing a human to guide every step.

Most interactions with AI follow a simple pattern: you ask something, the AI responds, you read the response. The AI is reactive. It waits for you, answers you, and stops. Agentic AI changes this into something closer to a colleague taking on a task: you describe what you want done, and the AI figures out the steps, executes them, deals with problems along the way, and comes back with a completed result.

To do this, an AI agent needs more than just language capability. It needs to be able to use tools - searching the web, running code, reading files, sending requests to other systems. It needs to be able to plan: break a goal into sub-tasks and decide which order to tackle them in. And it needs to be able to evaluate its own outputs - check whether what it just did worked, notice errors, and adjust its approach rather than blindly continuing.

The way most current agents work is by running the AI in a loop. The model takes a goal, decides on the next action to take, executes that action using a tool, gets the result back, decides on the next action, and so on - until it reaches a stopping point or determines the task is complete. This loop can run for many steps, sometimes dozens or hundreds, depending on the complexity of the task.

The potential productivity gains from this are large. Tasks that require a human to spend hours moving between different applications, gathering information, running analyses, and synthesising results could potentially be delegated to an agent. The AI does not get bored, does not lose track, and can run many tasks in parallel.

The risks are also real and taken seriously. An agent with access to real tools can take real actions - send emails, modify files, make purchases, delete data. Mistakes or misunderstandings get amplified across many steps. This is why the most thoughtfully built agent systems include checkpoints where the AI pauses and asks a human to confirm before taking irreversible actions, and why careful control over what tools an agent can access is considered essential.

Analogy

The difference between a consultant and a project manager. A consultant gives you advice when you ask. A project manager takes ownership of the outcome: plans the work, assigns tasks, follows up on progress, handles problems, and delivers a finished result. You check in on the project manager, not on every individual step. Agentic AI is the project manager model.

Real-world example

Devin, an AI system built by Cognition and released in 2024, is an agentic system for software engineering. You describe what you need built, and Devin writes the code, runs tests, reads error messages, fixes the bugs it finds, and iterates until the task is done - completing real engineering work with far less step-by-step human guidance than conventional coding assistants.

Why it matters

Agents are the next major shift in how AI creates value. The transition from AI as a question-answering tool to AI as an autonomous executor of complex tasks is where most enterprise AI investment is now directed. It is also where the most active safety research is happening - because agents that can act in the world need to be trusted in ways that conversational AI does not.

In the news

Related concepts

Hallucination Inference MCP (Model Context Protocol)Prompt Engineering

← Back to concepts