Asif Razzaq writes that Google Research has open-sourced Harness², a framework that improves agent performance by recursively refining the surrounding harness—prompts, tools, memory, and orchestration—while keeping model weights frozen. It operates without ground-truth labels, using a Propose, Probe, Compose cycle to test candidate harness edits against observed behavior.
- Weaker executors gained more; GLM-5 beat base Gemini 3.5 Flash on WorkBuddy-Bench after one refinement step
- Parallel recursion at k=5 lifted Gemini under OpenCode to 80.8 on WorkBuddy
- The framework extends the 7-file split from earlier Agentic Harness Engineering work by 1 component
- Requires Linux, Python 3.12+, uv, and Vertex AI for model-backed runs
- Weaker executors gained more; GLM-5 beat base Gemini 3.5 Flash on WorkBuddy-Bench after one refinement step
- Parallel recursion at k=5 lifted Gemini under OpenCode to 80.8 on WorkBuddy
- The framework extends the 7-file split from earlier Agentic Harness Engineering work by 1 component
- Requires Linux, Python 3.12+, uv, and Vertex AI for model-backed runs
Chejian Xu and co-authors write about Evo2, a framework for weight-harness co-evolution in self-evolving agents. Existing methods typically improve either model weights or the agent harness separately, overlooking the relationship between the two. Evo2 proposes a dual-role pipeline where a single policy serves as both a Proposer (writing harness edits) and a Reflector (diagnosing their effects). The training pipeline EvoTIDE recovers usable trajectories through edit extraction and prefix continuation, combining supervised fine-tuning with interleaved GRPO.
- Achieves 82.1% on Polyglot and 48.3% on SWE-bench Verified, outperforming the strongest baselines (62.6% and 41.0% respectively)
- Evolved harnesses transfer to different tasks without further evolution and rank first in cross-benchmark evaluations
- The framework extends to proprietary models, improving a GPT-5-mini agent's performance from 24.4% to 48.3%
- Achieves 82.1% on Polyglot and 48.3% on SWE-bench Verified, outperforming the strongest baselines (62.6% and 41.0% respectively)
- Evolved harnesses transfer to different tasks without further evolution and rank first in cross-benchmark evaluations
- The framework extends to proprietary models, improving a GPT-5-mini agent's performance from 24.4% to 48.3%
Paul Sawers writes that Docsy, the Google-created documentation theme for the Hugo static site generator, is moving to the Linux Foundation as AI agents become primary consumers of technical documentation.
- Docsy now generates Markdown copies and llms.txt files for LLM indexing, with an upcoming "AF" score to rate how readable docs are for agents.
- The project has been used by around 2,200 open source projects, including Kubernetes and OpenTelemetry.
- Docsy now generates Markdown copies and llms.txt files for LLM indexing, with an upcoming "AF" score to rate how readable docs are for agents.
- The project has been used by around 2,200 open source projects, including Kubernetes and OpenTelemetry.
Michał Januszewski and Viren Jain write that researchers have released a complete wiring diagram of the male fruit fly's brain and central nervous system, marking it as the largest brain map to date. In collaboration with HHMI Janelia, this project utilized computing and advanced techniques to create an annotated cellular-scale map containing over 166,000 neurons and 125 million synaptic connections. This foundational resource aims to help scientists understand how nervous systems function and may eventually assist in understanding complex human brain processes like memory or disease.
- The connectome includes the ventral nerve cord, which functions similarly to a spinal cord.
- Researchers are using similar mapping techniques on vertebrates, including fish and mice.
- This map complements previously released female fruit fly connectomes, allowing for studies of sexual dimorphism in behavior like courtship or aggression.
- The connectome includes the ventral nerve cord, which functions similarly to a spinal cord.
- Researchers are using similar mapping techniques on vertebrates, including fish and mice.
- This map complements previously released female fruit fly connectomes, allowing for studies of sexual dimorphism in behavior like courtship or aggression.
Mantis is a modular, stack-agnostic toolkit of security review skills designed for AI coding agents to autonomously discover, reproduce, and patch vulnerabilities. It provides a sequential pipeline—ranging from architecture synthesis and threat modeling to automated vulnerability research and remediation—to enable continuous and iterative defensive security reviews in an isolated environment.
- Designed for use with various coding agent frameworks like Gemini CLI or Antigravity SDK.
- Emphasizes the need for execution within sandboxed environments (e.g., gVisor, Docker) due to the risk of AI-generated code.
- Includes a "Snapshot Model" that allows for pinned, authoritative security passes on living codebases via non-destructive synchronization.
- Features an extensive set of specialized skills including `mantis-researcher` for flaw sweeping and `mantis-patch` for automated remediation.
- Designed for use with various coding agent frameworks like Gemini CLI or Antigravity SDK.
- Emphasizes the need for execution within sandboxed environments (e.g., gVisor, Docker) due to the risk of AI-generated code.
- Includes a "Snapshot Model" that allows for pinned, authoritative security passes on living codebases via non-destructive synchronization.
- Features an extensive set of specialized skills including `mantis-researcher` for flaw sweeping and `mantis-patch` for automated remediation.
Chao Zhang and colleagues study the design space of proactive AI agents that provide higher-level cognitive support during writing, moving beyond simple textual assistance like autocomplete. Through a technology probe deployed with 16 participants, researchers found that users prefer to configure custom partners by setting specific roles and proactivity levels in advance. The findings indicate that such tools can be used for both idea generation and self-monitoring when interventions are presented through lightweight visual representations and non-directive framing to minimize intrusiveness.
- Participants planned their AI support prospectively rather than reacting to real-time interruptions.
- Suggestions served a dual purpose of sparking new ideas and assisting in the self-monitoring process.
- The study emphasizes that rhetorical framing is as critical to user experience as the timing of an intervention.
- Participants planned their AI support prospectively rather than reacting to real-time interruptions.
- Suggestions served a dual purpose of sparking new ideas and assisting in the self-monitoring process.
- The study emphasizes that rhetorical framing is as critical to user experience as the timing of an intervention.
Google's research shows that LLMs experience difficulty recalling facts when questions reverse the usual subject/object entity order.
Roger Montti writes that Google has released version 0.2 of the Open Knowledge Format, incorporating five trust signals to facilitate data verification for consumers. These updates cover provenance, generation and verification status, freshness via expiration dates, lifecycle stages, and computational attestation to address questions regarding a concept's origin, accuracy, and timeliness. By providing raw metadata rather than subjective credibility scores, Google enables users or agents to dynamically evaluate information based on objective signals such as authorship and usage counts.
- Attested Computation is a new type used to confirm that values are calculated using sanctioned methods instead of improvised logic.
- The "sources" field provides objective data points like author and last modified date rather than an inherent score.
- Verification can be performed by humans or machine agents, allowing users to filter content based on verification level.
- Attested Computation is a new type used to confirm that values are calculated using sanctioned methods instead of improvised logic.
- The "sources" field provides objective data points like author and last modified date rather than an inherent score.
- Verification can be performed by humans or machine agents, allowing users to filter content based on verification level.
Google is rolling out its latest model, Gemini 3.5 Flash-Lite, within Google Search to support agentic search experiences and potentially enhance features like AI Overviews and AI Mode. This new model focuses on providing low latency and high throughput for workflows such as document processing and multi-step subagent tasks.
* Improved instruction following and better understanding of user intent
* Significant performance gains in coding (Terminal-Bench) and real-world task execution
* Optimized for agentic systems through built-in computer use tools and high-volume scalability
* Improved instruction following and better understanding of user intent
* Significant performance gains in coding (Terminal-Bench) and real-world task execution
* Optimized for agentic systems through built-in computer use tools and high-volume scalability
This guide provides a comprehensive walkthrough on using Google's Gemma 4 model to build autonomous AI agents through tool calling. It explores how this feature enables models to move beyond simple text generation by interacting with external APIs and systems via structured function calls.
Key topics covered in the article include:
- The mechanics of the tool calling loop, from reasoning and selection to execution and final response.
- Setting up a Python development environment using Hugging Face and necessary libraries like transformers and torch.
- Defining JSON schemas for tools to ensure precise model understanding.
- Implementing a full agent workflow by parsing function call responses and executing Python functions.
- A practical end-to-end demonstration of building a weather lookup agent.
- Managing multi-turn conversations through state management and conversation history.
- Best practices for production deployment, including argument validation, execution timeouts, and logging.
Key topics covered in the article include:
- The mechanics of the tool calling loop, from reasoning and selection to execution and final response.
- Setting up a Python development environment using Hugging Face and necessary libraries like transformers and torch.
- Defining JSON schemas for tools to ensure precise model understanding.
- Implementing a full agent workflow by parsing function call responses and executing Python functions.
- A practical end-to-end demonstration of building a weather lookup agent.
- Managing multi-turn conversations through state management and conversation history.
- Best practices for production deployment, including argument validation, execution timeouts, and logging.
