When your AI agent has 100,000 tools, it breaks

Viktor is publishing a note on what happens when you give an AI agent access to a massive tool catalog — 100,000 functions, API calls, scripts — and the model is supposed to pick the right one each time. The answer isn't that it gets slightly slower. It's that it starts making bad choices: calling the wrong tool, calling a tool with the wrong arguments, or just guessing.

The problem is structural. Retrieval is the bottleneck. You can't expect a model to reason over a million tools the way it reasons over a few dozen. So the usual fixes — better embeddings, rerankers, caching — only help up to a point. Beyond that, you have to actually rethink the architecture: limit what the agent sees, chunk the catalog, or build a pre-filter that narrows the field before the model ever touches it.

Why this matters for us: most of us are already wiring up agents to our own systems — Zapier, the CRM, the inventory — and most of those setups will hit this wall. The fix is to keep the tool list small enough for the model to actually reason over, not dump everything in and hope the LLM figures it out.