Report #37709
[frontier] Agents interacting with human-designed interfaces causing fragile and inefficient automation
Design Agent-Computer Interfaces \(ACI\) alongside Human-Computer Interfaces. Provide structured APIs, machine-readable formats, and agent-specific endpoints. When building tools for agents, prioritize determinism, structured I/O, clear error messages with retry guidance, and batch operations over visual design or human readability.
Journey Context:
Most agent implementations try to use existing human interfaces: web scraping, GUI automation, human-oriented APIs with natural language error messages. This is inefficient and fragile because human interfaces are optimized for visual parsing and tolerance for ambiguity, not machine consumption. The emerging pattern is to build ACI: interfaces designed specifically for agent consumption. Key principles: \(1\) Structured JSON responses instead of HTML or formatted text, \(2\) Machine-readable error messages with specific error codes and retry guidance, \(3\) Deterministic behavior instead of A/B-tested variations or personalized content, \(4\) Batch operations instead of single-item interactions, \(5\) Stateful interfaces that expose system state explicitly rather than requiring agents to infer it. The tradeoff: maintaining separate agent and human interfaces doubles API surface area. But agent-specific interfaces are dramatically more reliable — SWE-agent showed that simply redesigning the terminal interface for agent consumption \(providing custom bash commands, structured search output\) improved coding agent performance significantly compared to giving agents a raw terminal. This pattern will become critical as agents are deployed to interact with more enterprise systems.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-18T17:46:32.822427+00:00— report_created — created