Foundational design principles for Agent Experience (AX)

DIRECT Principles

Building autonomous AI-friendly software

D

Deterministic

idempotent predictable

Agents trust the system to behave the same way each time, making safe recovery from errors natural

  • Same input → same output, every time
  • Failed operations can be safely retried
  • Enables autonomous error recovery
I

Introspectable

discoverable structured

Agents explore capabilities and understand data structures without requiring external docs or humans

  • Self-describing APIs and data schemas
  • No need to read documentation
  • Agents discover capabilities on their own
R

Recoverable

retryable debuggable

Agents can diagnose what went wrong and take corrective action without human intervention

  • Clear error messages with context
  • Agents understand what failed and why
  • Actionable recovery paths available
E

Explicit

unambiguous precise

Agents operate within clear boundaries and constraints so they act with confidence

  • Unambiguous parameters and behavior
  • Clear input validation and constraints
  • Reduces agent hesitation and errors
C

Consistent

uniform learnable

Agents build mental models once and apply them everywhere, scaling their effectiveness

  • Uniform patterns across all APIs
  • Agents learn rules once, apply everywhere
  • Reduces cognitive load exponentially
T

Testable

dry-run sandboxed

Agents can try actions safely before committing, learning what works without causing harm

  • Dry-run or sandbox modes available
  • Safe experimentation without consequences
  • Enables agents to learn by trying
Created by John Wang — Docs · GitHub