OpenAI made Codex smart enough that the bottleneck moved. Most people haven’t noticed where it went.

Source: Nate’s Newsletter Date: 2026-05-09 URL: https://natesnewsletter.substack.com/p/codex-plugins-bottleneck-moved

Summary

Nate’s Newsletter argues that Codex’s capability gains (Terminal-Bench 2.0: 75.1% → 82.7%) have shifted the real constraint from model quality to workflow setup cost. Users still manually re-establish context, standards, and tool access each session — spending real effort just to make the agent ready to act. Codex plugins are the proposed fix: packaged skill bundles that carry tool integrations, deterministic validation, team standards, and institutional knowledge across sessions without re-loading.

Implications

• Feeds the session continuity thread directly: the “40% of your AI time gone to setup” framing quantifies the cost of stateless agents, which is the same problem driving persistent memory and workflow tooling across CLI agents.
• The competitive edge the piece identifies — professionals who package reusable workflow infrastructure — maps exactly to what the Workflow tool (Claude Code v2.1.147) and Codex goals mode are each trying to own.
• Plugin/workflow portability is becoming a product surface in its own right; watch for ecosystem lock-in through workflow formats, not just model quality.