Early-Signal Detector

method · first seen 4 Oct 2026

tool calling

Also: tool calls, agent tool use, function calling, LLM tool calling, LLM-based tool use, tool-calling agents, tool use by LLM agents

A mechanism letting language models invoke external functions or APIs with structured arguments.

Tool calling, also known as function calling, lets a language model emit structured requests (a function name plus JSON-like arguments) that a host program executes, feeding results back to the model. It has been a standard feature of major LLM APIs and agent frameworks since 2023, so the concept itself is not new. The sources show ordinary background use: a paper on topology-aware task planning over tool dependency graphs, a paper cataloguing agent-harness invariants, DeepSeek Harness plugins, an Android agent, and Chinese notes for a Stanford agents course. The late weekly spike is driven largely by a burst of near-identical 'ai--manus' repositories from throwaway accounts, which looks like star farming rather than real adoption.

Tool calling is basic infrastructure for LLM agents, so current research is moving toward harness design, planning over tool graphs, and safety boundaries rather than the mechanism itself. The apparent surge here reflects duplicate repositories and broad usage, not a new idea gaining ground.

Analyst note (established), 17 Nov 2026.

Mentions per week

027719 May11 Aug10 NovPapersCodeDiscussion

PapersCodeDiscussion

Rank 5 on 14 Nov 2026 · 346 mentions in 28 days · 253 independent groups · papers → code → discussion · first seen 6 weeks ago · acceleration +2.50

Earliest sources

Latest sources

Mentioned together with

sandboxed code executionagent sandboxingManusmodel context protocolretrieval-augmented generationsandboxingprompt injectionspeculative decodingBerkeley Function Calling Leaderboardfault injection