Chatbots answer. Agents act. In 2026 the useful ones share a pattern: plan → tool call → check → retry.
What’s shipping
- IDE / repo agents — Claude Code, Cursor agents, Copilot Agent Mode. Best ROI for most builders.
- Computer-use / browser agents — strong for research and ops glue; still need supervision on money-moving steps.
- Multi-agent frameworks — LangGraph, CrewAI, AutoGen-style stacks. Fine for orchestration; don’t confuse a demo graph with a production SLA.
The model underneath still dominates outcomes. Today’s AA LLM leaders (Opus 5, GPT-5.6 Sol) change how far an agent gets before it stalls.
Rules of thumb
- One agent, one clear goal, observable tools
- Prefer products with logs and diffs over black-box “auto employees”
- Re-bench when the index moves — wrappers age slower than models