AI Briefing#agent-harness
The latest AI and developer news about #agent-harness, with the original source and a short summary.
The latest AI and developer news about #agent-harness, with the original source and a short summary.
InfraRSI achieved a geometric mean speedup of 1.83x on H100/H200 by simultaneously exploring configurations and code.
IBM's open-source CUGA agent harness achieved an 81.1% task completion rate on the AppWorld benchmark, significantly outperforming Hermes and Deep Agents using the same GLM-5.3-flash model.
The new system compiles JavaScript into executable cells as the model generates them, reducing tool steps by 43% and cached input tokens by 39%.
Factory releases Droid 1.0, featuring Programmatic Tool Calling (PTC) and a 37% reduction in first-request input tokens.
The harness, which powers GitHub Copilot and VS Code, supports 20+ frontier models and is built by the GitHub engineering team.
Anthropic's new plugin system allows developers to customize how Claude Code interrupts and interacts with them, shifting the agent harness from machine-facing to human-facing.
Ramp developed Laminar to unify its agentic products on a model-agnostic, cloud-native stack, replacing fragmented vendor solutions and open-source frameworks.
Cache misses caused by harness modifications like timestamp changes or tool list updates increase energy consumption by up to 2.5x compared to fully cached requests.
Integrating the hosted Jev decision model with Qwen 3.8 27B reduced execution API costs by 26.6% while maintaining 100% task accuracy on 360 SOP-Bench cases.
Using NVIDIA NeMo Relay to trace Hermes Agent before and after harness changes revealed that the success rate increase for Qwen3 Coder 30B was partly due to eliminating provider errors, while average LLM calls and duration increased.
The new middleware reduced median cost per coding task by 64% in Open SWE experiments without impacting quality.
The update introduces official Windows and macOS desktop clients, experimental async question handling, and a new permission-diagnosis skill for Windows Sandbox.
Raven is a modular agent harness designed for recursive self-improvement, featuring built-in agents and an Evolver tool that refines the harness based on benchmark evaluations.
Introduced seven latest AI research results, including improving task success rates from 23.3% to 44.3% through agent harness distillation.
According to a Mozilla report, the performance gap between open and closed models has narrowed to an average of about 4 months, with competition shifting from the models themselves to harnesses.
OpenAI and Anthropic are offering agent loops as managed APIs, pursuing a strategy of integrated optimization of models and harnesses.
Salesforce researchers proposed an On-Policy Correction pipeline to overcome the limitations of imitation learning for weaker models and facilitate the co-evolution of harnesses and models.
It emphasized that unavoidable complexity should be absorbed at the core abstraction layer rather than offloaded to extensions or users.
Based on experience building agent harnesses, this article provides a detailed analysis of the internal working mechanisms of Skills and MCP, along with harness optimization techniques.
TrueForge has released an open-source agent execution environment that integrates LLM tool calling, sandboxing, and session management.
The EvoUndo framework, designed to recover from self-correction failures in LLM agents, has been published in a paper.
DeepSeek-AI and Peking University have released Cordis, a formal theory-based meta-framework for runtime plugin removal and dependency management.
As LLMs internalize harness capabilities, the remaining harness evolves into an interface for human attention rather than for the model.
TrueFoundry has released TrueForge, an open-source harness that integrates the agent execution loop from model calls to sandboxes and UI.