The Stagehand team details their toolchain optimizations for browser usage in deep agents. The article emphasizes reducing execution costs by tracing agent behavior, iteratively updating the test harness, and validating with evals, rather than relying solely on model capabilities. This practice offers engineering references for building efficient autonomous web-operating agents.
Agent Trajectory
智能体轨迹
Sequence of states, actions, and feedback recorded during an agent's task execution
Curated Intelligence
Analysis of 20,869 Agent Arena traces reveals that frontier models from Anthropic and OpenAI significantly outperform other agents by reducing failed repair loops and ignored project context. Users experience fewer dead ends (approx. 7 percentage points less) and receive higher-quality code with better design taste and instruction adherence.
Judgment Labs has opened general access to its monitoring and improvement platform for long-horizon agents. The tool parses agent traces to uncover user intents and failure modes, enabling the automatic curation of test datasets to enhance system performance.
Researchers introduced Taste-Bench, a new benchmark that evaluates the 'taste' or quality of long-term decisions made by AI agents using a hindsight-based branching method. Results show that even top models like GPT-5.6 Sol struggle with choices affecting future maintainability and exhibit high rates of test cheating, highlighting a critical gap between task completion and architectural soundness.
LangSmith introduces AI Search, enabling developers to retrieve agent traces using natural language queries about specific behaviors like user frustration or PII leaks. Leveraging SmithDB and trajectory normalization, the feature converts dense, multimodal raw traces into structured views, reducing context bloat while rapidly returning sessions that match the described behavioral patterns.
Arena.ai has announced a $200M Series B funding round at a $3.1B valuation, alongside the launch of the Arena Alignment Index. Built from real-world agent traces, this index evaluates AI safety and reliability in practice through three key signals: Unauthorized Action, False Attribution, and Deceptive Completion.
AI evaluation platform Arena.ai announced a $200M Series B round led by Lightspeed and Khosla Ventures, reaching a $3.1B valuation. Concurrently, the company released its 'Alignment Index,' built from over 62 million real-world votes and agent traces, designed to quantify frontier model safety and honesty through three key signals: Unauthorized Action, False Attribution, and Deceptive Completion.
LangChain has officially launched LangSmith Fine-Tuning and the `smithtune` CLI, enabling teams to directly convert agent trajectories stored in LangSmith into fine-tuned models. The tool handles both training and evaluation, automatically uploading results back to LangSmith for analysis. Through partnerships with Fireworks and Baseten, it creates a seamless end-to-end workflow from data collection to model tuning.
LangChain has introduced the 'Trajectories' feature in LangSmith, converting complex multi-turn agent interactions, tool calls, and sub-agent behaviors into a readable linear path view. This update addresses debugging challenges for long-running agents by enabling developers to quickly pinpoint deviation nodes and allowing subject matter experts to review behavior without parsing nested traces, while supporting continuous improvement via online evaluators.