APIs built with human developers as primary users are increasingly unsuited for AI agents, which don’t share human intuitions or access to “tribal knowledge.” Legacy APIs often rely on undocumented behaviors, inconsistencies, or implicit expectations that humans can compensate for—but AI agents cannot.

---

Problems with Legacy APIs

• Assumed Knowledge & Implicit Contracts: Many APIs implicitly assume that callers know about quirks, conventions, or undocumented behavior. AI agents lack that shared context.
• Poor Error Handling / Ambiguous Behavior: Legacy APIs may have inconsistent or vague error messages or responses. Such ambiguity confuses AI agents.
• Non-machine-friendly Interfaces: Some APIs are not structured for automated discovery, versioning, or predictable behavior—things AI agents rely on.

---

What Needs to Change

• Clear, precise contracts: APIs should have well-specified, machine-readable contracts, with minimal ambiguity.
• Better versioning and backward compatibility: Changes to APIs should be managed so agents don’t break (or be clearly told when they might).
• Improved observability and error information: Agents need to be able to detect, trace, and recover from failures or unexpected outputs.
• Cultural/organizational shift: API design and maintenance processes need to evolve, with AI consumption in mind — not just human developers.

A very detailed analysis of API and their limitations with AI Agents: https://thenewstack.io/why-your-legacy-apis-are-a-roadblock-for-ai-agents/