AI agent harnesses

Google Research has open-sourced RRSI, a framework allowing AI agents to refine their own evaluation harness while guarding against overfitting. Announced via MarkTechPost, the release lets developers inspect and build on the underlying code. The announcement is drawing attention from AI practitioners interested in agent self-improvement methods that remain reliable rather than gaming their own benchmarks.

Univer is an open-source TypeScript project from dream-num that bills itself as an 'Office Harness for AI Agents'. It provides a single runtime combining spreadsheets, documents, slides, canvas, relational tables, and PDF handling. The repository is trending on GitHub, and the framing suggests developers are interested in giving AI agents tools to create and manipulate office-style documents. Beyond the project's own description, there is little discussion in the available evidence explaining what users are saying about it.

A JavaScript project called ECC, published by developer affaan-m, is gaining traction among developers. It bills itself as an agent harness performance optimization system, offering skills, instincts, memory, security, and research-first development workflows for AI coding tools including Claude Code, Codex, Opencode, Cursor and others. It is currently trending as one of the most viewed new repositories, reflecting continued interest in tooling that improves how AI coding agents perform.

A GitHub project by michael-denyer, pstack-claude, adapts Poteto's pstack — rigorous agent workflows built on Cursor primitives — for other AI coding harnesses, including Claude Code, Codex, Copilot, Pi, OpenCode, Gemini and Prime Agent. It is gaining traction among developers who want structured, disciplined agent workflows outside Cursor, reflecting growing interest in porting agent tooling across platforms.

Orbi, an AI coding agent, reportedly takes a GitHub issue and returns a reviewed, merged pull request, with one agent writing the fix and another handling review. A follow-up post examining the system's harness asks what its reviewer missed, suggesting developers are scrutinising how reliable the automated pipeline is. The discussion is drawing interest from the open-source and AI engineering communities.

A developer has released Television, an open source graphical interface for AI agent harnesses, sharing it with the technology community. The tool, available at television.run, aims to give users a visual way to manage and interact with coding or task agents rather than working purely from the command line. Early reactions are modest so far, but tools that simplify agent workflows are drawing attention as AI agents become more widely used in software development.

The MLC team has released the TIRx Harness, an open-source compiler harness designed for agentic GPU programming, where AI agents write and optimize GPU code. The project, detailed in a blog post by researcher Jinhong Yi, combines compiler infrastructure with automated agent workflows. It is drawing attention from developers interested in the intersection of machine learning systems, compilers, and autonomous code generation.

A new guide has been released explaining how to train AI agents across multiple harness environments, so the same agent can be developed and tested on different tooling frameworks. It targets developers working with agentic AI systems who want their models to behave consistently regardless of which harness runs them. Response has been moderate so far, with discussion focused on practical implementation details.

DeepSeek has released version 0.2 of its Harness, the open-source agent framework, introducing official desktop applications for the first time. The update means users can now run the AI agent harness through dedicated desktop apps rather than relying solely on command-line or manual setups. Coverage so far is limited to tech outlets, with little community reaction recorded yet.

A widely shared essay argues that in the era of AI agents, the harness—the scaffolding of tools, prompts, evaluation and workflow code wrapped around a model—is where a company's actual value sits, not the underlying model itself. As models become commoditised and interchangeable, the author contends the harness is the durable product, and effectively the company's true identity.

AWS has launched a locally run, open source tool for keeping tabs on AI agent harnesses, the software frameworks that let autonomous AI systems take actions. The offering gives developers a way to monitor and constrain agent behaviour on their own infrastructure rather than relying on hosted services. It reflects growing demand for guardrails as companies deploy agentic AI in production.

A developer has released Raven on GitHub, describing it as 'the harness of harnesses, built for RSI' — recursive self-improvement in AI systems. The tool, from EverMind-AI, is presented as a layer that coordinates multiple evaluation or agent harnesses. It drew quick attention on Hacker News, where commenters were probing how it works and what it actually automates.

A new article on gihyo.jp, the Japanese developer site run by Impress and technical publisher Gijutsu Hyoronsha, explains agentic coding and harness engineering: the tooling, guardrails and design practices used to maximise how far AI coding agents can work autonomously. The piece is being shared in Japanese developer circles, with readers tagging it alongside discussions of agentic AI and software design.

A New Stack commentary argues that Kubernetes' history of breaking free from monolithic designs offers a cautionary lesson for builders of AI agent harnesses. The piece suggests teams designing agent frameworks should avoid tightly coupled, monolithic architectures, drawing parallels with how container orchestration evolved toward modularity. Discussion is centered on software architecture practices for the AI era.

AI startup Manus has released version 2.0 of its AI agent platform. The update introduces the Cascade agent harness, which cut token usage by 23.2% and costs by 32% in testing. A new app called Cue gives each agent its own email address, phone number and digital wallet, and a China-specific version is planned.