This past week, OpenAI shipped Record & Replay for Codex.
Show Codex a workflow once. It turns the demonstration into an inspectable, editable skill that can guide similar work later through Computer Use, browser actions, and installed plugins.
Anthropic moved in the same direction through a different architecture: agent teams, composable skills, and nested subagents in Claude Code.
Two stacks. Same week. Same shift.
From research conversation to shipped primitive
The thing that was a research conversation last month and a procurement decision last week is becoming a shipped primitive.
In the OpenAI version:
- record a workflow
- extract the repeatable procedure
- save it as a callable skill
In the Anthropic version:
- describe a skill
- load it into specialized agents
- compose agents and skills into a larger workflow
One learns from demonstration. The other composes explicit procedures. Both move the unit of work up a level.
Workflow intelligence as a product
You are no longer only wiring tools to prompts. You are encoding how work is actually done and making that procedure callable by agents.
That is workflow intelligence as a product.
And it sharpens the ownership question.
Who owns the recording, the corrections, the exceptions, and the accumulated procedure the agent inherits?
The vendor? Or the builder?
What the Cursor deal may have been pricing
Perhaps that is part of what the Cursor deal was pricing.
Not only the editor. The workflow layer around it.
If memory was part of what last week's procurement conversation was about, workflow intelligence is what this week shipped. The two layers are arriving as products around the same time, and the ownership question follows them both.
Originally posted on LinkedIn on June 22, 2026 — Edward Izgorodin.
Related: Memory is becoming a procurement decision (last week's piece on the same arc).
More research on AI agent memory and workflow intelligence at mnemoverse.com/docs/library. I build Mnemoverse, open-source persistent memory for AI agents.
Top comments (0)