An agent is an AI system that autonomously plans, takes actions through tools and iterates toward a goal. Where a chatbot is a one-shot question-and-answer, an agent decides what to do next, does it, observes the result, and continues. The agent loop — perceive, plan, act, observe — has become the dominant pattern for any AI feature that does more than generate text.
A typical agent in 2026 has these components:
- An LLM (usually Claude Sonnet 4 or GPT-5 for strong reasoning, sometimes a smaller specialised model for cost) that drives the loop.
- A set of tools the LLM can call: search, database query, file read/write, browser automation, code execution, custom APIs.
- Memory — short-term conversation history and sometimes longer-term persisted memory across sessions.
- A planner — explicit or implicit, the mechanism that breaks a goal into steps.
- Guardrails — input filters, output filters, action approvals for high-stakes operations.
The flavours of agents that have stabilised by 2026:
- Task agents — given a discrete goal ("research these 10 companies and summarise findings"), they execute and return.
- Coding agents — Claude Code, Cursor Composer, GitHub Copilot Workspace, Cognition Devon. Edit code, run tests, iterate.
- Computer-use agents — Anthropic Computer Use, OpenAI Operator. Control a sandboxed desktop or browser as if they were a human.
- Customer-facing agents — voice agents, support chat agents, sales agents that interact with end users autonomously.
- Multi-agent systems — multiple specialised agents (researcher, writer, critic) coordinating on a task.
The frameworks people actually use:
- Direct API calls with tool calling — for most production cases, the provider's native API is enough.
- Anthropic Agents SDK / OpenAI Assistants API — first-party orchestration with persistence and threading.
- LangChain / LangGraph — older but widely adopted; LangGraph specifically for graph-based agent workflows.
- CrewAI / AutoGen — multi-agent orchestration with role definitions.
- Pydantic AI / Instructor — type-safe agents in Python.
- Mastra / Vercel AI SDK — TypeScript-native agent frameworks.
The hard problems remain hard. Long-horizon planning, recovering from errors, knowing when to stop, avoiding cost explosions and prompt injection are all still active challenges. The 2026 mature pattern is: keep agents focused on narrow well-defined tasks, give them strong evaluation harnesses, sandbox aggressively, and put humans in the loop for any irreversible action. Agents that try to do everything are unreliable; agents that do one thing extremely well are transforming entire job categories.
