In what reads like coordinated timing, OpenAI and Anthropic both shipped workflow intelligence primitives this past week—OpenAI with Record & Replay for Codex, Anthropic through agent teams, composable skills, and nested subagents in Claude Code. The convergence isn't coincidence. It's the next phase of the AI stack materializing faster than most builders expected.

OpenAI's Approach: Learn From Demonstration

Show Codex a workflow once. It extracts the repeatable procedure, saves it as a callable skill, and can replay that procedure later through Computer Use, browser actions, and installed plugins. The key word is "inspectable"—you're not just getting black-box automation. You can see what the model learned, edit it, and iterate. Record & Replay turns live demonstration into persistent institutional knowledge.

Anthropic's Approach: Compose Explicit Procedures

Anthropic took a different path but landed in the same territory. Rather than learning from demonstrations, Claude Code now lets you describe skills directly, load them into specialized agents, and compose multiple agents and skills into larger workflows. The model doesn't watch and infer—it receives explicit procedures and executes them through structured subagent orchestration.

What This Means for Builders

The implications are significant: you're no longer just wiring tools to prompts. You're encoding how work is actually done—decision trees, exception handling, the accumulated tribal knowledge of your team's processes—and making that procedure callable by agents at scale. That shift from "prompt engineering" to "workflow engineering" represents a fundamental change in what it means to build with AI.

The Ownership Question Nobody Is Asking (Yet)

But here's what's nagging at me: who owns the recording? Who controls the corrections, the exceptions, and the accumulated procedure that your agent inherits? If you're building on someone else's platform, are you building your workflow intelligence or just renting it? This question is about to become a procurement conversation, if it isn't already.

The Cursor Deal in Context

That brings me back to the Cursor acquisition chatter. Perhaps what was being priced wasn't just the editor—it was the workflow layer around it. Memory (last week's primitive) and workflow intelligence (this week's primitive) are arriving as products simultaneously. If you've been watching AI agent infrastructure, you know these layers compound each other.

Key Takeaways

  • Record & Replay for Codex turns demonstrations into callable, editable skills
  • Claude Code's agent composition achieves similar goals through explicit procedures
  • Both approaches move "the unit of work" up a level—from prompts to processes
  • Ownership of workflow intelligence is an open question with serious procurement implications

The Bottom Line

The AI stack is solidifying fast. If you're not thinking about how your team's workflows become agent-native procedures, you're leaving the next layer of leverage on the table. Watch this space—workflow intelligence primitives are shipping now, and whoever controls them controls the automation moat.