Skip to main content
human input
LLM reasoning
tool call
tool result
autonomous
output
Technical AI
What Are AI Agents?
AI agents can browse the web, write and execute code, book meetings, and manage files — all without a human in every loop. Here's how they work and what to watch out for.
The shift happening now
$47B
Projected AI agent market by 2030 — growing at 44% CAGR from $5.1B in 2024 [MarketsandMarkets]
700
AI agents replacing customer service roles at Klarna in 2024 — handling 2.3M conversations in the first month [Klarna]
14%
Of SWE-bench coding tasks solved autonomously by Devin — the first AI software engineer to reach double-digit benchmark performance [Cognition AI]
A standard LLM is reactive — you ask it something, it responds, interaction complete. An AI agent is different: it has a goal, access to tools, and a loop where it can plan, act, observe results, and adjust its approach — all without waiting for you to tell it what to do next.
The shift from chatbot to agent is the shift from "AI that answers questions" to "AI that completes tasks." It's a meaningful difference. An agent can be given a high-level objective like "research this competitor, summarise their pricing, and draft a competitive positioning doc" and handle the entire sequence autonomously.
The four components
An AI agent needs four capabilities that a standard chatbot lacks. Remove any one of them and it's no longer truly agentic.
Planning
Breaks a high-level goal into sub-tasks. Decides what to do first, second, third.
Tool use
Can call external tools: web search, code execution, file system, APIs, browsers.
Memory
Stores task progress and intermediate results. Short-term (context) or long-term (vector DB).
Action loop
Observe results, reason about them, decide next action — repeatedly until goal is complete.
The ReAct loop
Most AI agents use some version of the ReAct (Reason + Act) pattern. Here's what one iteration looks like.
GOAL
Receive task
"Research the top 5 CRM tools, compare pricing, and write a summary table."
THINK
Reason about approach
The agent plans: "I need to search for each CRM, visit pricing pages, extract key numbers, then format a comparison."
ACT
Use a tool
Calls web_search("Salesforce CRM pricing 2026"). Gets results. Calls browser_navigate("salesforce.com/pricing").
OBSERVE
Read tool output
Receives pricing page content. Extracts: Starter $25/seat, Pro $75/seat, Enterprise custom. Stores in memory.
LOOP
Repeat for each CRM
Continues the same cycle for HubSpot, Zoho, Pipedrive, Monday CRM. Five loops later, it has all the data it needs.
DONE
Generate final output
Synthesises all gathered data into a formatted comparison table. Returns result to user.
Real examples
OpenAI Operator
OpenAI — launched Jan 2025
Browses the web and completes tasks autonomously: book restaurants, fill forms, research and purchase products. Uses a browser-based action space.
Devin
Cognition AI
Full software engineering agent. Plans and writes complete codebases, runs tests, debugs failures, deploys to production. First agent to pass SWE-bench at meaningful rates.
Claude Computer Use
Anthropic
Controls a desktop computer — moves mouse, clicks, types, navigates applications. Can complete multi-step workflows across any desktop software.
Microsoft Copilot Agents
Microsoft
Business process agents in Microsoft 365. HR agents answer employee questions. Sales agents update Dynamics CRM. IT agents resolve tickets autonomously.
AutoGPT
Open-source
Early open-source agent that spawns sub-agents, assigns them tasks, and coordinates their outputs. Demonstrated the potential before commercial products matured.
Klarna AI Agent
Klarna
Customer service agent handling refunds, disputes, payment questions, and order management. Replaced 700 human equivalent roles, handling 2.3M conversations in first month.
Risks before you deploy
Error compounding
Each agent action builds on previous ones. A wrong assumption in step 2 can compound into a completely wrong output by step 10. Agents need checkpoints and the ability to halt when confidence is low.
Prompt injection
Malicious content in web pages or documents can hijack an agent's behaviour. If your agent browses external websites, adversarial instructions embedded in those pages can cause unintended actions.
Irreversible actions
Agents with access to email, files, or financial systems can take actions that are hard or impossible to undo. Production agents need permission tiers — read access before write access, confirmation before delete.
Cost runaway
An agent loop that fails to complete can run thousands of LLM calls before timing out. Always set hard limits on token usage and number of iterations before deploying any autonomous agent in production.
The principle to follow
Start agents with minimal permissions. Give read access before write access. Give reversible actions before irreversible ones. Build in human checkpoints for anything consequential. An agent that asks for confirmation before deleting files is vastly preferable to one that silently gets things wrong.
FAQ
What's the difference between an AI agent and a chatbot?
A chatbot responds to inputs. An agent pursues goals autonomously. A chatbot waits for you to say "search for X." An agent, given the goal "research competitors", decides to search for X, Y, and Z, opens the relevant pages, extracts the data, and produces a report — without being told each individual step.
Are AI agents safe to use?
With appropriate constraints, yes. The risks are real but manageable. Limit what tools an agent can access, require confirmation for irreversible actions, set iteration limits, log all actions for review. Treat a new AI agent the way you'd treat a new junior employee — capable, but needing oversight until trust is established through performance.
What frameworks exist for building agents?
LangChain and LlamaIndex offer agent frameworks for Python. OpenAI has the Assistants API with tool use. Anthropic has Claude with tool use and computer use capabilities. Microsoft's AutoGen supports multi-agent coordination. CrewAI is popular for role-based agent teams. For production deployments, managed services from AWS, Azure, and GCP are increasingly available.
Sources
[MarketsandMarkets] MarketsandMarkets — AI Agent Market Report 2024
[Klarna] Klarna press release — "Klarna AI assistant handles two-thirds of customer service chats" (Feb 2024)
[Cognition] Cognition AI — Devin SWE-bench results (2024)
[Yao] Yao et al. — "ReAct: Synergizing Reasoning and Acting in Language Models" (2022)
Get AI insights every week
The AI Briefing covers what actually matters in AI — no hype, no jargon, just what you need to stay ahead.
Subscribe free