Specialized agents

Code agents (Cursor / Claude Code / Devin), browser / computer-use agents (Operator, Anthropic Computer Use), research agents, data agents — patterns and limits.

By 2026, several agent categories have matured into product surfaces with their own UX patterns: code agents in IDEs (Cursor, Claude Code, Aider, Cline, Devin), browser / computer-use agents that drive a screen (Anthropic Computer Use, OpenAI Operator), research agents (Perplexity, Deep Research), data agents (Hex, Julius). Each has lessons that transfer to other agent applications.

This module is being expanded.

Coming in the next revision:

Code agents. What makes them work — file-aware context, tool sets (read, edit, run tests, run shell), planning at the multi-file level. The Cursor-vs-Aider-vs-Claude-Code comparison.
Browser / computer-use agents. The mechanics — screenshots, mouse/keyboard simulation, vision models reading the screen. Failure modes: brittle on UI changes, latency-sensitive, slow.
Research agents. Multi-step web search + synthesis. The Deep Research pattern: dozens of searches before producing one answer. Cost economics.
Data agents. SQL generation, chart creation, iterative refinement. Hex/Julius patterns. Why “natural language to SQL” requires more than just an LLM.
Customer-support agents. Intercom Fin, Sierra, Decagon. Production guardrails, conversation handoff to human.
Patterns that transfer. Tool sandboxing, file-context windowing, “show your work” UX, undo-ability.

Next: Module 09 — Evaluation.