AI agents sometimes pick the wrong tool or the wrong partner — do the two main connection standards fail differently?

This explores how finding the right capability and picking the right tool break down differently in MCP and A2A. The corpus can only answer half of that: it has evidence about MCP and nothing on A2A, so it can't support a real side-by-side comparison. What it does have is a sharp picture of how MCP fails, which is the half of the comparison you'd need first.

The clearest MCP evidence comes from a production study. Wiring MCP into an agentic workflow caused non-deterministic failures, traced to ambiguous tool selection and parameter inference. The model saw several plausible tools, or plausible ways to fill in arguments, and sometimes chose differently on the same input. The fix was to drop the protocol layer for explicit direct function calls and give each agent a single tool, which restored determinism. The same study found that 85% of the 306 practitioners surveyed build custom agents rather than use frameworks (Why do protocol-based tool integrations fail in production workflows?). Here the failure is not that the tool can't be found. Too much choice is left to the model at call time.

The discovery side fails for a different reason. MCP-Zero argues that single-round semantic matching, where a retriever picks tools by similarity to the user's request, breaks on the gap between how people talk (colloquial) and how tools are described (formal). Its answer is to let the model itself emit structured tool requests, refining what it needs as its reasoning unfolds across domains, instead of a retriever choosing for it (Can models decide better than retrievers which tools to use?). So MCP has two separate failure points: discovery (vocabulary mismatch) and selection (ambiguity among candidates). Fixes for one don't fix the other.

Two notes from outside the tool-use literature suggest why discovery is hard. General retrieval research finds that embeddings measure association rather than task relevance, and that this is an architectural limit, not something tuning fixes (Where do retrieval systems fail and why?). Separately, letting an agent search raw text with exact-match commands recovers precision where embeddings blur similar entities together (Can direct corpus search beat embedding-based retrieval?). Neither note is about MCP or A2A, so treat this as a hypothesis: a tool-discovery step built on similarity search may inherit the same blind spots, and exact-name lookup might sidestep some of them.

The A2A half is missing. A2A is meant to connect agents to other agents, so its discovery problem would be finding a peer agent's capabilities, not a tool's. The corpus has no note testing whether that fails the same way, worse, or differently. The closest hint is that the production study's single-tool-per-agent design shrinks the choice each agent faces, but it doesn't measure delegation between agents. If you want the comparison, this is a real gap in the collection.