Tool Use

Tool Use lets an AI system call external functions, APIs, browsers, shells, or databases and act on the results, instead of only generating text. The model emits a structured call, the runtime executes it, and the output is fed back for the next step.

It is what connects a language model to the real world: fetching live data, running code, or changing state. Clear tool schemas and descriptions matter as much as the model, because the model selects tools from their definitions alone.

For builders it is the foundation of Agents and MCP integrations. The risks are safety and reliability: a tool that can act needs permission boundaries, validation, and error handling, since a hallucinated or malformed call can cause real side effects.

Topic: Agents Related: Agents MCP Context Engineering

Recent Updates

• 2026-08-30: monty-go: Safe Execution of LLM-Generated Python Code in Pure Go (breakingnewsofficial) · github.com — fugue-labs Anthropic · Cloudflare · Hugging Face · Pydantic
• 2026-08-31: Agents for Everyone: A Workshop Framework for Building Agentic AI Capabilities in a Distributed Curation Community (breakingnewsofficial) · arxiv.org — Anthropic Gene Ontology Consortium · Claude
• 2026-08-31: Prompt Injection Attack Hijacks Claude Code Opus 5 Auto Mode (breakingnewsofficial) · embracethered.com — Anthropic Trajectory Labs Boris Cherny · Opus 5
• 2026-09-01: Keenable SELECT: SQL-based agent for web research with semantic extraction (breakingnewsofficial) · keenableai.github.io
• 2026-09-02: Skill Following: Evaluating Actual Skill Use in Retrieval-Enabled LLM Agents (breakingnewsofficial) · arxiv.org
• 2026-09-02: Long-Horizon State Tracking in LLMs: Executing MD5 through a Deep Sequence of Dependent Tool Calls (breakingnewsofficial) · arxiv.org — GPT-OSS 120B
• 2026-09-02: Trajectory-Judge: Outcome-Only Evaluation Misses 55% of Silent Agent Faults (breakingnewsofficial) · arxiv.org
• 2026-09-03: CivBench: A Long-Horizon Benchmark for Tool-Mediated Agents in Civilization VI (breakingnewsofficial) · arxiv.org
• 2026-09-03: LLM-as-a-Judge Is Not an Oracle: Why Self-Improving Agents Need Deterministic Guardrails (breakingnewsofficial) · arxiv.org
• 2026-09-03: OpenAI releases GPT-6 Astra claiming major advances in agentic software engineering and computer automation (breakingnewsofficial) · qbitai.com — OpenAI Legora Playco Greg Brockman · GPT-6 Astra GPT-6 Astra Pro · GPT-5.6 Sol · Fable 5.1 · Claude Opus 5 · Gemini 3.8 Flash
• 2026-09-04: Interface-Induced Trajectory Censoring (breakingnewsofficial) · arxiv.org — Qwen2.5-coder · Llama-3.1-8B
• 2026-09-04: Coding agents don’t always choose the most precise tool: how retrieval interface design affects agent routing (breakingnewsofficial) · agentconnect.md — Opus 4.8 · Sonnet 4.6 · Haiku 4.5
• 2026-09-06: Pigeon: Capability-Based Authorization for AI Sub-Agents (breakingnewsofficial) · github.com — pigeonlabsHQ
• 2026-09-08: aispace: Agent-Optimized File Sharing with Expiring Links and Local Encryption (breakingnewsofficial) · github.com — aispace-sh GitHub · OpenAI · Anthropic
• 2026-09-09: SAP: State-Guided Data Synthesis with Argument Provenance for Multi-Turn Tool Use (breakingnewsofficial) · arxiv.org — SAP-4B
• 2026-09-09: DAREBench: Deployment-Aware and Reliable Evaluation of Models as Agents (breakingnewsofficial) · arxiv.org
• 2026-09-09: AgentBrew: Offline Tool-Use Agent Learning from Raw Real-World Trajectories (breakingnewsofficial) · arxiv.org — Qwen3-32B · Qwen3-235B
• 2026-09-09: Beyond Prompts: Measuring and Optimizing LLM Tool-Agent Harnesses (breakingnewsofficial) · arxiv.org
• 2026-09-09: MERIT: Cost-Aware Evaluation of Memory in Tool-Using LLM Agents (breakingnewsofficial) · arxiv.org — OpenAI · Anthropic · GPT-4.1-mini · GPT-4.1 · Claude Haiku 4.5 · Claude Sonnet 5
• 2026-09-09: XunFei Spark X2.5: End-to-End Agentic Task Delivery from Creative Prompts to Production Code (breakingnewsofficial) · qbitai.com — XunFei OpenAI · Anthropic · NVIDIA · Hugging Face · DeepSeek Spark X2.5 · Claude
• 2026-09-09: OtoDock: Self-hosted agent platform with persistent memory and team coordination (breakingnewsofficial) · github.com — Anthropic · OpenAI · Twilio
• 2026-09-10: AgentAudit: An Open, Extensible Framework for Full-Lifecycle Trust Evaluation of AI Agents (breakingnewsofficial) · arxiv.org — OpenAI · Anthropic · Google · Meta Sarvam AI · GPT-5 · Claude Sonnet 5 Sarvam 105B · Llama-3.3-70b · Gemini 2.5 Flash
• 2026-09-10: State-Path Tool Menus: Learning Execution Routes to Improve Agent Tool Selection (breakingnewsofficial) · arxiv.org
• 2026-09-10: HybridDeepResearch Benchmark Reveals AI Agents Struggle With Web and Database Reasoning (breakingnewsofficial) · arxiv.org — Snowflake · OpenAI · Anthropic · Hugging Face · GLM-5.2 · Claude Sonnet 4.6 · GPT-5
• 2026-09-10: How AI agents now compress multi-day prototyping into minutes with Codex and tool orchestration (breakingnewsofficial) · t.me

FAQ

What is Tool Use?

Tool Use lets AI systems call external functions, APIs, browsers, shells, or databases instead of only generating text. GROUNDING tracks tool schemas, agent loops, safety boundaries, and reliability patterns.

Which topic does Tool Use belong to?

On the GROUNDING radar, Tool Use is grouped under the Agents topic.

Which concepts are related to Tool Use?

Related concepts tracked by the radar include Agents, MCP, Context Engineering.