Make the wrong thing impossible
summary
Argues that APIs must make the wrong thing impossible by giving agents choices instead of asking them to generate values: use post-to-slot capability URLs so repeated submissions return the same successful response, shift decisions to runtime, and expose HATEOAS links instead of human-readable manipulable URLs. Enforce closed schema enums (no optionality that explodes document space) so compilers/tests catch errors the agent can fix immediately. Demonstrates with temp-file capability and idempotent slot examples and reports that the approach survives agent-driven refactors.
AI-written summary · about
chapters (15)
- 0:00Introduction
- 1:10Identity key problem in APIs
- 4:19Agents struggle with identity keys
- 8:59Moving errors to compile time
- 10:03Understanding HATEOAS
- 12:47HATEOAS prevents invalid actions
- 15:56Human-readable URLs anti-pattern
- 17:01Solving identity key problem with HATEOAS
- 18:27Capability models and security
- 20:33Capability models in practice
- 21:39Validating agent-generated code with schemas
- 22:42Closed versus open sets in APIs
- 24:04Handling errors with HATEOAS links
- 25:48Selection beats generation for agents
- 26:09Future of API design with agentic engineering
related talks
35mUnleash Your Inner API Architect: Build Agents That Deliver
18mBuild a terrible API for people you hate
34mDesign Principles: API vs MCP
56mWhen Unstoppable Frontier Agents Meet Immovable Attack Vectors
29mThe missing layer: What your APIs need before agents can use them
24mThe API Economy's Next Chapter: AI Agents as First-Class Clients
44mDesigning Hypermedia APIs
01h 36mThe Most Important API Design Guideline - No, It's Not That One