Autonomous Agents
Also known as: AI agent, LLM agent, agentic AI, software agent
Autonomous agents are AI systems that can pursue a goal over a series of steps, making decisions, using tools, and adapting their approach based on intermediate results -- without requiring a human to approve every individual action.
They differ from chatbots (which respond to single messages) and copilots (which suggest actions for humans to approve) by taking sequences of actions autonomously to complete a task. Modern AI agents built on large language models can browse the web, write and execute code, send emails, query databases, and call external APIs to accomplish multi-step goals.
Autonomous Agents, explained at three levels
At home
An autonomous agent is an AI that can do a sequence of things on its own to achieve a goal. If you ask an AI chatbot to 'book me a flight to Toronto', it tells you how to do it but cannot do it for you. An autonomous agent can actually do it: search for available flights, compare prices, fill in your traveller details, and complete the booking -- taking each step itself without you having to approve every click. You give it the goal and it figures out and executes the steps.
At work
Autonomous agents are used in business workflows to automate sequences of tasks that would previously require a human to coordinate multiple systems. An agent might monitor a company's support inbox, classify incoming requests, retrieve relevant knowledge base articles, draft responses, and route escalations to human agents -- handling routine queries end-to-end. The key difference from traditional automation (like RPA) is that agents can handle variation and judgment calls that rigid scripts cannot.
At home
AI agents use a planning and acting loop. They start with a goal, break it down into steps, execute the first step using available tools, observe the result, and decide what to do next based on what they learn. The tools available to an agent might include web search, code execution, database queries, API calls, file reading/writing, and email sending. The agent reasons in natural language between each tool use, which allows it to handle unexpected results and change its plan mid-task.
At work
Autonomous agents introduce new categories of risk compared to simpler AI systems. Mistakes compound: unlike a chatbot where each response is independent, an agent's early errors can cascade through a multi-step task before a human notices. Scope creep is a risk: agents given broad tool access might take actions outside the intended scope. Irreversible actions (sending emails, executing transactions, deleting files) cannot be undone if the agent makes an error. Agent deployments in production environments require strict tool permission boundaries and human approval gates for high-consequence actions.
At home
The 'agentic' design pattern involves three key elements: a language model as the reasoning engine, a set of tools the agent can invoke, and a memory system that allows the agent to retain context across steps. More complex multi-agent architectures use several specialized agents (a planning agent, a research agent, an execution agent, a verification agent) that communicate with each other, with a coordinator agent directing the workflow. This pattern improves performance on complex tasks by decomposing them into sub-tasks handled by specialized sub-agents.
At work
From a practitioner standpoint, the fundamental challenge in autonomous agent deployment is trust and control. The agent's autonomy is only valuable if it produces correct results; incorrect autonomous actions may be harder to detect and correct than incorrect chatbot responses. Production agent deployments typically use three mechanisms: a strict tool permission model (the agent can only use tools explicitly granted for the current task), a human approval gate for any action above a specified risk threshold (sending external communications, modifying production data, spending money), and a complete audit log of every action and decision.
Real-world examples
At home
Ryan asked his AI assistant to 'research the best noise-cancelling headphones under £200, create a comparison table, and save it to my notes'. The agent searched three review sites, extracted product data, formatted a comparison table in his note-taking app, and sent him a notification -- completing a task that would have taken him 45 minutes in about three minutes.
At work
A legal firm deployed an AI agent to handle initial contract review. Given a new vendor contract, the agent: extracted key clauses using document analysis tools, compared them against the firm's standard template using a comparison tool, flagged deviations, and drafted a summary memo for the reviewing lawyer. The lawyer still reviewed every contract, but the prep work that previously took two hours was completed in five minutes.
How it compares
| Feature | Autonomous Agent | AI Copilot | Chatbot |
|---|---|---|---|
| Autonomy level | High -- executes multi-step tasks independently | Medium -- suggests actions; human approves each | Low -- responds to single messages; takes no actions |
| Tool use | Yes -- can invoke APIs, browse web, run code, write files | Sometimes -- may have limited tool integrations | Rarely -- mostly generates text |
| Error risk | Higher -- mistakes can compound across steps | Lower -- human reviews each step | Lowest -- outputs are informational, not actions |
| Appropriate for | Well-defined multi-step tasks with clear goals and boundaries | Productivity augmentation where human judgment is needed | Q&A, information retrieval, conversation |
| Human oversight | Approval gates for high-risk actions; full audit log | Per-action approval by design | User reviews and acts on responses |
Agents are more autonomous and capable than copilots or chatbots but require more careful oversight design. Copilots keep humans in the loop at every step. Chatbots are conversational only. The right choice depends on the task complexity, acceptable error rate, and consequences of mistakes.
For analysts
In BA artefacts, autonomous agent requirements need to specify boundaries at least as carefully as capabilities. Define: which tools the agent has access to; which actions require human approval before execution; what constitutes a task completion vs. escalation to a human; and what information is logged for audit. When writing acceptance criteria, specify the error handling path: 'If the agent is unable to complete a step in the workflow, it must halt, log the incomplete action, and notify the responsible human operator rather than attempting to proceed or find an alternative path.' Agents should not have write access to production systems without explicit, bounded permissions.
History & origin
The concept of autonomous agents in AI research dates to the 1980s with software agents -- programs that could act on behalf of users to filter email, negotiate prices, or monitor conditions. The term gained renewed currency with the rise of LLM-based agents from 2023. AutoGPT (April 2023) was an early widely publicized experiment showing a GPT-4-based system attempting to complete open-ended tasks autonomously. The field moved quickly from experiment to product, with LangChain, OpenAI Assistants API, Anthropic's Claude with tool use, and Microsoft Copilot Studio all providing frameworks for building agents. Agentic AI became a major industry theme in 2024-2025.