TL;DR Teaching Qwen3 models (1.7B-8B) tool selection over 10,000+ tools with RL beats prompting the 402B Llama-4-Maverick on the same job. Title: Training Small Language Models for Transferable Tool Selection URL: Key points ・🧩 Plan-Retrieve-Select: split the instruction into substeps, retrieve candidates, pick one tool per substep ・🧹 Noisy ToolBench G3 gold labels are filtered with a teacher model down to 20,093 examples ・🎯 SFT then GRPO: rank-weighted F1 for Plan, exact match for Select ・📈 ToolBench G3 success: 41.5% at 8B vs. 30.6% for the best prompted Maverick ・🔁 Swap BM25 for the dense retriever COLT and accuracy improves with no retraining ・🧪 A learned selector adds about 13 points over a BM25 top-1 rule ・⚠️ Planning transfers poorly to TaskBench, where a large prompted model still wins A good example of why a small, purpose-trained model can beat a giant prompted one on cost, speed and accuracy. #LLMAgents# #ToolUse#

# Useful but Little-Known Features of Claude Agent SDK 🌍 Losing track of complex multi-step tasks? Structured task tracking makes progress visible! Claude Agent SDK includes built-in task management for organized progress tracking during agent execution. 📌 Title: Structured Task (Todo) Tracking 🔗 URL: 🧩 Overview Four tools — `TaskCreate`, `TaskUpdate`, `TaskGet`, and `TaskList` — enable structured task management during agent execution. These task-tracking tools are available by default only on Claude 3.x models, Opus 4 through 4.7, Sonnet 4 through 4.6, and Haiku 4.5 (the default set in Claude Code v2.1.268 and later, bundled by the TypeScript SDK from v0.3.268); on every other model you must opt the session in. The agent automatically registers multi-step work as tasks and manages their lifecycle: `pending` → `in_progress` → `completed`. Unlike the legacy `TodoWrite` which overwrote the entire list on every call, Task tools operate on individual items, enabling fine-grained progress tracking. 🛠 How to Use ```python # Python - Monitor TaskCreate/TaskUpdate in the stream from claude_agent_sdk import query, ClaudeAgentOptions, AssistantMessage, ToolUseBlock async for message in query( prompt="Optimize my React app performance", # Keeps the Task tools on models where Claude Code otherwise doesn't provide them options=ClaudeAgentOptions(env={"CLAUDE_CODE_ENABLE_TODO_TOOLS": "1"}), ): if not isinstance(message, AssistantMessage): continue for block in message.content: if not isinstance(block, ToolUseBlock): continue if == "TaskCreate": print(f"+ {block.input['subject']}") elif == "TaskUpdate" and block.input.get("status"): # The streamed input is the raw shape the model emitted (may be id / task_id) task_id = ( block.input.get("taskId") or block.input.get("id") or block.input.get("task_id") ) print(f" {task_id} -> {block.input['status']}") ``` ```typescript // TypeScript - Display task creation and updates in real time import { query } from "@anthropic-ai/claude-agent-sdk"; for await (const message of query({ prompt: "Optimize my React app performance", // Keeps the Task tools on models where Claude Code otherwise doesn't provide them options: { env: { ...process.env, CLAUDE_CODE_ENABLE_TODO_TOOLS: "1" } }, })) { if (message.type !== "assistant") continue; for (const block of message.message.content) { if (block.type !== "tool_use") continue; if ( === "TaskCreate") { console.log(`+ ${block.input.subject}`); } else if ( === "TaskUpdate") { // The streamed input is the raw shape the model emitted (may be id / task_id) const taskId = block.input.taskId ?? ?? block.input.task_id; if (taskId && block.input.status) console.log(` ${taskId} -> ${block.input.status}`); } } } ``` 🏗 Integration into Production Systems - Reflect task progress in UI progress bars or checklists to give users visibility into agent work - Capture `taskId` from `TaskCreate` tool results and maintain a map for a complete task overview - Monitor `TaskList` tool results for snapshots to display on dashboards - Tasks are automatically created for 3+ step workflows, making them ideal for monitoring long-running operations 💡 Use Cases 📋 Share large-scale refactoring progress with the team in real time 🔄 Track completion status of each agent step in CI/CD pipelines 📊 Display which analyses are complete on a dashboard during multi-step code review ⚠️ Caveats - `TaskCreate` input does not contain `taskId`; retrieve it from the corresponding `tool_result` block as `{ task: { id, subject } }` - On models that don't have the tools by default, opt the session in by naming a tool in `allowed_tools`, listing the tools in the `tools` option, or setting `CLAUDE_CODE_ENABLE_TODO_TOOLS=1` in `env` — without opting in, no `tool_use` blocks for them appear in the stream. Conversely, `CLAUDE_CODE_ENABLE_TASKS=0` makes Claude Code provide the legacy `TodoWrite` instead of the four Task tools - Setting `status` to `"deleted"` removes a task ✨ Waiting for multi-step work without knowing what's happening is stressful. Use task tracking to make your agent's work fully transparent! #ClaudeAgentSDK# #AIAgent#

🔧 Turns out the efficient way to build tool-use training data is to write the answer first. Title: ToolGrad: Efficient tool-use dataset generation with textual "gradients" URL: Google Research's ToolGrad flips the usual order of tool-use dataset generation. Here are 3 highlights. 🔄 Answer-first, not query-first Instead of generating a user instruction first and searching depth-first for a solution, ToolGrad builds a verified API workflow first, then derives the matching user prompt, no trial-and-error exploration needed. 📝 "Gradients" made of text Instead of numeric gradients, natural-language feedback from API execution reports becomes a "textual gradient," driving four modules, Proposer, Executors, Selector, and Updater, that build the workflow step by step. 🏆 Matches top proprietary models Fine-tuned on a small dataset, ToolGrad-12B scores 83.1 on the Berkeley Function Calling Leaderboard, on par with Gemini 2.5 Pro and Claude 4.5 Opus, and ahead of GPT-5. The striking part: cheaper data generation translates directly into real, competitive agent performance. #AIAgents# #ToolUse#

ATLAS turns industrial tool-use evaluation into a dual-horizon diagnostic loop, then reuses calibrated signals for policy optimization validated on Meituan Xiaotuan production traffic. ATLAS: Dual-Horizon Diagnostic Evaluation for Industrial Tool-Use Agents Paper:

Hermes Wingtips #48#: platform toolsets your agent pays a fixed token overhead before it reads a single word of your message, and one of the biggest slice is not the prompt. every tool your platform loads ships its full schema with every call, used or not. if a session never touches the browser, image gen or TTS, `platform_toolsets` in config.yaml hands that platform a smaller set. `hermes prompt-size` shows you the before and after without spending a token.

OpenAI says it will pause training with tool use on its most capable models until the sandbox flaw is resolved. "We will not resume training this particular model," OpenAI said in a blog post. Read more:

Hermes Wingtips #79#: `agent[.]disabled_toolsets` Hermes Agent picks its tools per platform: the CLI, Telegram, Discord and the rest each keep their own list of toolsets, set in `hermes tools`. this key is one list above all of them. a toolset named here is removed everywhere at once, on the CLI and on every gateway platform. it is applied after the per-platform lists. `web` takes web_search and web_extract out of every session. `memory` removes the memory tools and also drops the memory guidance from the prompt.