AI DevelopmentSimon Willison

Use subagents and custom agents in Codex

Read original
ai-coding-toolscursor-vs-copilotautomation-stacks

“Subagents pattern is now widely supported across coding agents - from OpenAI Codex to Claude Code to Cursor, with custom agents definable as TOML files”

Key takeaways

  • OpenAI Codex launched subagents in general availability with default agents for 'explorer', 'worker', and 'default' - similar to Claude Code's implementation
  • Custom agents can be defined as TOML files in ~/.codex/agents/ with custom instructions and specific model assignments (including gpt-5.3-codex-spark for speed)
  • Subagents pattern has achieved cross-platform standardization - now supported by OpenAI Codex, Claude Code, Gemini CLI, Mistral Vibe, OpenCode, VS Code, and Cursor, signaling architectural convergence in AI coding tools
  • The pattern enables specialized agent orchestration (e.g., 'browser_debugger' reproduces issues, 'code_mapper' traces paths, 'ui_fixer' implements fixes) for complex debugging workflows
  • Platform convergence suggests subagents are becoming the standard architecture for agentic coding workflows, similar to how REST APIs became standard for web services

Why this matters for operators: Engineering teams evaluating AI coding assistants; understanding emerging standards in agentic workflows

I cover AI×GTM intelligence like this every Wednesday.

Get STEEPWORKS Weekly

More picks

Human-AI Intersectionr/artificial

Why Hasn’t AI Made Work Easier?

  • Large-scale study (164K workers, 180-day tracking) shows AI adoption doubled time spent on email/messaging/chat and increased business software use by 94%, but reduced focused work time by 9%
  • This represents a 'productivity paradox'—AI accelerates shallow, context-switching work while cannibalizing the deep work that drives actual value creation
  • Pattern repeats historical technology adoption cycles (email, mobile, video-conferencing) where efficiency tools paradoxically increased busyness without proportional output gains
ai-productivity-paradoxshallow-work-trapdeep-work-decline
Read original Full analysis →
Personal Productivity & AI-Augmented WorkTechCrunch AI

Cursor admits its new coding model was built on top of Moonshot AI’s Kimi

  • Cursor's new coding model is built on Chinese AI company Moonshot AI's Kimi foundation model
  • This represents a significant supply chain transparency issue in a widely-adopted developer tool
  • Geopolitical tensions around Chinese AI models create regulatory and compliance risk for enterprises using Cursor
ai-coding-toolscursor-vs-copilotregulatory-impact
Read original Full analysis →

This analysis was produced using the STEEPWORKS system — the same agents, skills, and knowledge architecture available in the GrowthOS package.