Tag archive
We ran an OpenAI API probe on three tool-misuse faults an order agent hits (schema drift, rate limits, fake-success replies) and recorded the recovery.
A practical, CI-runnable map of what changes when an LLM can use tools: tool invocation, connectors/OAuth, memory/RAG, browser/OS control, and tool supply chain.
Function Calling and the Architecture of Useful Agents
Most people who use AI every day are still treating it like a very sophisticated autocomplete. Type a...
Tool-using agents often act on what they think is true rather than what they've checked. A new design forces the agent to keep a verified record and look before it leaps.
A new paper pins the sudden collapse of multi-step tool-use training on runaway probabilities in a few control tokens, and shows that mixing in supervised examples stabilizes it.
Flask creator Armin Ronacher found that Anthropic's newest models, Opus 4.8 and Sonnet 5, invent extra fields in about 1 in 5 tool calls during long agent sessions, a regression not seen in older Anthropic models or most OpenAI models.
Our in-product chatbot doesn't just answer questions — it proposes concrete changes to a running app through a typed toolbox, the same function-calling idea behind MCP. Toolbar buttons, validations, computed columns, conditional styles, even SQL fragments. Every proposal is a chip with an Apply button. This post walks the catalogue, one tool at a time, with the prompt that triggers each.
SciAgentGYM benchmarks LLM agents across 1,780 scientific tools in Physics, Chemistry, Biology — even GPT-5 drops below 31% on hard multi-step tasks.
LLM 에이전트로 마케팅 리포트를 자동화할 때 어디까지가 진짜 쓸만하고 어디서 깨지는지, tool-use·환각 방지·실무 패턴을 마케터 시각으로 정리.
LLM이 광고 API·BigQuery·Slack을 직접 부르기 시작하면, 답변 품질보다 "어느 도구를 언제 부를지"가 운영 사고의 진앙이 됩니다. function calling의 한 줄 직관과 마케터가 점검할 5가지.
How Microsoft's ARTIST framework uses outcome-based RL to train LLMs that interleave tool calls inside reasoning chains — no step supervision required.