1. Hermes Agent Introduces Tool Search to Handle Large MCP Catalogs
Nous Research's Hermes Agent now includes a Tool Search feature designed to bypass context window limitations when integrating large Model Context Protocol (MCP) toolkits. By replacing full tool schemas with a progressive-disclosure layer, the agent matches queries against tool descriptions using the BM25 algorithm. Evaluated on Anthropic's models, the system dramatically improves task accuracy while keeping the tool catalog stateless to prevent synchronization drift.
- • Introduces three bridge tools—tool_search, tool_describe, and tool_call—to dynamically fetch schemas only when needed.
- • Leverages the BM25 retrieval algorithm for matching queries against tool metadata with a substring fallback.
- • Improves accuracy from 49% to 74% on Claude Opus 4 and 79.5% to 88.1% on Claude Opus 4.5 in Anthropic's evaluations.
- • Activates automatically when deferrable tool schemas consume over 10% of the active model's context.
- • Excludes core terminal and web search tools to keep them directly accessible, while rebuilding the stateless tool catalog on every turn.
Developers building with large Model Context Protocol (MCP) catalogs can now prevent context bloat while significantly increasing agent accuracy.
2. Text-to-Speech Leaderboard Ranks Top API and Open-Weight Models
Evaluating the top performers in the text-to-speech landscape is now easier with the latest Artificial Analysis Speech Arena benchmarks. The leaderboard showcases highly optimized options, ranging from ultra-low latency models like Cartesia's Sonic 3.5 to deeply controllable systems like ElevenLabs Eleven v3 and Gemini 3.1 Flash TTS which parse inline tags for pacing. For developers seeking open-weight alternatives, Fish Audio S2 Pro leads the rankings but mandates a commercial license for production deployment.
- • Gemini 3.1 Flash TTS, Realtime TTS-2, Sonic 3.5, and Realtime TTS 1.5 Max lead the Artificial Analysis Speech Arena as of May 30, 2026.
- • Cartesia's Sonic 3.5 uses a State Space Model architecture to achieve end-to-end latency of approximately 82 milliseconds.
- • Inworld AI's TTS-1.5 Mini targets real-time use cases with a P90 time-to-first-audio under 130 milliseconds.
- • ElevenLabs Eleven v3 and Gemini 3.1 Flash TTS support inline audio formatting tags for style and pacing control.
- • Fish Audio S2 Pro is the highest-ranked open-weight model but requires a commercial license for non-research use.
- • OpenAI's GPT-Realtime-2 introduces voice capabilities with GPT-5-class reasoning for live speech-to-speech interactions.
Developers integrating audio into their apps can compare real-world latency, reasoning, and licensing terms across leading commercial and open-weight TTS models.
3. AgentTrove Releases Streaming Pipeline for Agent SFT Datasets
Creating fine-tuning datasets from agentic behavior is simplified with AgentTrove's new Python pipeline. The open-source tool streams traces directly, letting developers filter for successful outcomes and parse tool usage without storing gigabytes of raw files. The extracted sequences can be exported immediately into the ShareGPT format, making it straightforward to feed into common training libraries like Axolotl or LLaMA-Factory.
- • AgentTrove contains 1.7 million agentic interaction traces accessible via Python-based streaming.
- • Includes a command-extraction utility that parses shell commands from assistant JSON outputs to evaluate tool use.
- • Employs a success-based filter to isolate traces marked as resolved, passed, or positively rewarded.
- • Exports filtered traces into a clean ShareGPT-style JSONL format compatible with Axolotl and LLaMA-Factory.
- • Utilizes pandas and matplotlib to summarize and visualize task sources, teacher models, and turn counts.
Developers can construct high-quality, success-filtered supervised fine-tuning datasets from complex agent interactions without downloading massive raw data files.