the agent surface map: pick the cell before the prompt
Most "agent" products fail by shipping the wrong cell's interface and credentials. Surface and autonomy choose the product before the model does.
Product @ Okta | Developer Experience & Tooling | Identity Platform Strategy
Most "agent" products fail by shipping the wrong cell's interface and credentials. Surface and autonomy choose the product before the model does.
I scanned 200 public SKILL.md files. 96% fail lint, and 95% of the valid ones omit the "Use when…" line that routes activation.
Agents retry, fail, and disappear without a complaint. Their behavior creates a product discipline I call agent experience.
Mintlify says AI agents account for 48% of docs traffic. A normal 404 gives them no useful recovery path.
An observation becomes an insight when you can explain the mechanism, predict a different outcome, and change a decision.
Claude Code nails developer experience. But as capabilities grow, so does complexity. The best tools disappear - they don't become things you have to manage.
Neutral observations on the Vercel vs Cloudflare race to win the edge and developer mindshare
AI speeds up execution. Letting it frame the problem before you do can weaken the judgment product teams are paid to develop.
Large B2B platforms own broad categories. Smaller teams can win workflows those platforms neglect.
Social feeds turn other people's milestones into instructions. You can spend years chasing goals that someone else chose.
A polished strategy document cannot replace contact with users, competitor tradeoffs, and command of the domain.
Thin API wrappers dump orchestration onto the model. Useful MCP servers encode a workflow, its domain rules, and its safety boundaries.
F1 teams prepare branches, triggers, and decision windows before the race. Product teams can use the same discipline under uncertainty.