February 27, 2026

TL;DR. Stop automating workflows designed for humans — redesign for outcomes and collapse the tiers entirely.

Takeaways

• The default enterprise AI playbook (map workflow → automate steps) just makes broken processes faster, not better.
• Most workflows exist because of human limitations (training, quality control, error containment) — AI removes those constraints.
• The real opening for AI startups: outcome-based automation, not workflow automation. Three support tiers should collapse into one direct resolution.

TL;DR. Designing agent tools means watching how the model actually behaves — Claude Code's team iterated through three failed elicitation approaches before landing on the AskUserQuestion tool.

Takeaways

• Tool design is shaped to the agent's abilities, like picking paper vs. calculator vs. computer for a math problem — match the tool to what the model can actually wield.
• First attempts at structured questioning failed: cramming questions into ExitPlanTool confused Claude; loose markdown formats were unreliable. A dedicated tool with a blocking modal finally worked.
• "Even the best designed tool doesn't work if Claude doesn't understand how to call it" — model preference matters as much as interface design.
• What works for one model may not work for the next; tools like Todos that helped earlier Claude versions may need rethinking as capabilities evolve.