Thariq Shihipar, speaking independently after leaving Anthropic, uses a fighting-game demo to explain human-centered agent workflows at CODE Summit 2026.

Stop treating today’s model habits as permanent

  • Shihipar once expected Bash, React and diffusion-generated art to dominate their respective workflows.
  • He now prefers MCP, the standard for connecting agents to tools, for some interactions and sees capable models using multiple languages and creative applications.
  • Skills, code execution and progressively revealed tool information offer additional ways to structure an agent’s work.

Industrial revolution and Renaissance

  • Software factories make implementation more repeatable and plentiful, in his framing.
  • Creativity, utility and careful design can become more valuable as basic production gets easier.
  • He wants developers to pursue both reliable production and distinctive software.

The jagged frontier and smooth interior

  • The “jagged frontier” describes the uneven edge of what a model can achieve.
  • The “smooth interior” describes the breadth of methods it handles once a task is comfortably within reach.
  • A mockup might arrive as HTML, React or a Figma file; a video might come through direct application control.
  • His imagined model that needs little checking is a future ideal, not evidence that current security review is unnecessary.

When the agent builds something you no longer understand

  • In his game, “throw a rock” leaves open speed, charge, size and whether the projectile passes through a wall.
  • The agent fills those gaps, leaving him unsure how his own character works.
  • Prompting, leaving and returning also breaks his concentration and slows his progress.

Give people an editable layer

  • He separates engine internals from a readable folder of character abilities and rules.
  • Comments, lint rules and meaningful units make that layer understandable enough to inspect and change directly.
  • The result resembles a custom scripting language: an executable specification for the decisions he wants to own.
  • In the demo, he changes rock size and tries Cursor’s tab completion for an action, while Claude still builds more complex primitives.

Design the working experience

  • Readability, fast iteration, new languages, new primitives and older human-computer interaction research all become design opportunities.
  • Shihipar presents this as an early experiment and admits he has yet to settle on a final workflow.
  • Caveat: the talk offers a design argument and a small live example, not measured productivity gains or proof that model limitations have disappeared.

“We have this like the AI can work in the way that we want, right?” — Thariq Shihipar, 12:03