An empirical log-based study reconstructs a two-week collaboration in which a user and AI coding agents built a procedural strawberry animation: the user sent 3,958 prompts, 74% of them while an agent was working, with at least 24 false status reports, 37 user interventions against invisible or unrequested work, and 17 misrouted messages. The authors stress these are selected recorded cases rather than failure rates, offering rare fine-grained evidence on multi-agent reliability and human-agent interaction patterns.
Tool Calling
工具调用
The ability of models to invoke external functions or tools, where multi-step orchestration and batched calls are becoming key t
Curated Intelligence
Microsoft unveiled its in-house coding models MAI Code 2 and MAI Code 2 Flash at Ignite 2026, now live in GitHub Copilot. Microsoft claims Code 2 averages about 98% of GPT-6.1 Sol's performance across four coding benchmarks at $0.309 per task, roughly 5% of Claude Opus 5.5's cost, while the lightweight Flash scores a 61.0% average pass rate on DeepSWE 1.1, Terminal-Bench 2.1 and SWE Atlas, beating GPT-6 Luna and Claude Haiku 5.5.
Artificial Analysis built VIALS-AA on Handshake AI's VIALS benchmark to test visual reasoning over life-science artifacts such as gel blots, plasmid maps and phylogenies. Claude Fable 5.2 leads at 67%, ahead of Gemini 4 Carbon (64%) and GPT-6 Astra (63%). The more notable signal is tool-use strategy: Fable averages roughly 37 tool calls per task with heavy batching, versus about 8 for Astra, suggesting reasoning models are shifting toward multi-step tool orchestration rather than single-shot answers.
A joint team at Shanghai Jiao Tong University has built AutoDNA (Siyuan-1), an AI-native scientific laboratory that autonomously performs molecular biology experiments. A multi-agent architecture decomposes tasks into literature search, protocol design, code generation, hardware execution and iterative optimization, while an 'atomic service' abstraction exposes heterogeneous instruments to the model through natural-language descriptions and unified code interfaces. The system achieves micron-level positioning and microliter-scale liquid handling, and was published in Nature with a homepage feature, including an end-to-end DNA storage workflow demo.
Tool-calling agents need thousands of verifiable tasks for RL training, yet benchmarks offer only dozens to hundreds and LLM-synthesized tasks often lack reliable ground truth. This NeurIPS 2026 work inverts the pipeline: explore real MCP servers first, then derive tasks from observed outcomes, yielding 5,144 verified tasks across 240 servers and 993 tools, plus an offline-reproducible retrieval-augmented tool-call simulator.
A developer reports that with two connectors pointing at the same MCP server, ChatGPT silently routed a follow-up tool call—carrying a one-time confirmation code from the first connector's widget—through the second connector without notifying the user. The issue touches session isolation and authorization boundaries in agent tool calling, but the reporter declined to file it because the bug bounty required biometric and ID data upfront; no official response yet.
Artificial Analysis launched Endpoint Accuracy Index v1.1, benchmarking over 100 serverless inference endpoints and finding that nearly one in four scores below a self-hosted reference deployment of official weights. The index scores endpoints equally on tool calling, scientific reasoning and long-context recall, helping developers choose providers on accuracy rather than price and speed alone.
AICon Beijing will run Dec 18-19 with 13 tracks spanning agent engineering, inference optimization, and agent evaluation and safety. Tencent engineer Chen Siying will present an AI-native enterprise dev toolchain in the Agentic Software Engineering track, covering tool protocolization, agent runtime engineering, context engineering platforms, and a friction-signal feedback loop. The piece is primarily an agenda announcement without concrete technical parameters or verifiable results.
Tencent announced Hy4, its latest open-source large language model, claiming performance that ranks among the world's leading open models. Hy4 adds visual understanding and computer-use capabilities for the first time, alongside gains in instruction following, long-context comprehension and long-horizon agentic tasks. Model weights and system cards are slated for release in the coming days, with Tencent emphasizing responsible open-source development around cybersecurity and biological safety.
The author describes a general technique for reverse-engineering agent products: instead of directly asking for system prompts or skill definitions, first have the agent load and execute the target skill, then ask it to reproduce the same inputs and outputs or draft a technical plan from that primed context. The approach works across products like Claude Design and Claude Duet, though complex skills may be reproduced incompletely and require follow-up questions.