The real gold rush in agentic AI
The most valuable discoveries are often tiny workflow nuggets found in real work, then written down so they compound.
The real gold rush in agentic AI is not finding one perfect model or prompt. It is discovering tiny workflow nuggets in real work, then capturing them so the next task starts from a better place.
A useful handoff format. A short markdown rule. A review checklist. A better way to describe visual intent. Individually, these improvements look small. Together, they change the quality and predictability of the whole development process.
A concrete example
Codex can be very strong at implementation, but a scratch user interface may still drift toward boxy cards, excessive glow, or weak hierarchy. Telling an agent to “make it better” rarely produces a reusable improvement.
A more reliable workflow is:
- Take a screenshot of the current interface.
- Use image generation to explore a stronger reference design.
- Refine the direction: remove generic AI glow, tighten hierarchy, and clarify what should lead.
- Give the reference back to the coding agent for implementation.
- Encode the lesson as an
AGENTS.mdrule or another durable project instruction.
The first four steps improve one screen. The fifth improves every later screen.
Best practices are being discovered in real time
Agentic development is young enough that many of its best practices are not in textbooks. They emerge while building: where context gets lost, where reviews become expensive, which decisions should be made before code, and which recurring mistakes deserve a permanent rule.
The teams that benefit most will not merely use agents more often. They will notice these small discoveries, write them down, and turn individual judgment into shared infrastructure.
That is the gold: not a dramatic trick, but a growing collection of workflow improvements that compound.
Adapted from a post originally published on LinkedIn.