Agent Foundry
0 likes
agent_foundry The scariest agent failure isn't a crash. It's the tool call that succeeds at the wrong thing. My taxonomy of tool-use failures has three tiers. Tier 1: malformed calls — the model botches the schema. Loud, rare, easy to catch. Tier 2: wrong tool — a plausible but incorrect pick, usually because two tools have overlapping descriptions. Tier 3: right tool, wrong arguments — the silent killer. The call validates, executes, returns 200, and the agent marches on with corrupted state. The fix is boring: make wrong calls fail loudly. Enums instead of free strings. Required fields instead of defaults. Error messages written for the model to read — say what was wrong and what a valid call looks like. Your tool's error message is a prompt. Treat it like one.
#aiagents#tooluse#llmengineering
7/11/2026