AI Agents

MaxKernel combines large language models with compiler feedback and hardware profiling to automate TPU kernel development. Its system supports human-guided collaboration, autonomous optimization, and graph-based exploration.

Iris-mini and Iris-pro are trained on multi-hop search tasks reconstructed from web link structures. The work combines supervised fine-tuning, online reinforcement learning, and explicit context management to improve long-horizon browsing.

Terminal-Universe proposes reconstructing executable workspaces from existing terminal-agent trajectories instead of building every environment from scratch. The recovered environments can support original-task replay, new task synthesis, cross-repository queries, and multi-turn interactions.

As terminal agents improve, tasks generated from scratch can become too easy to provide useful feedback. A new environment evolution approach raises task difficulty across generations and supplies sustained signals for long-horizon reinforcement learning.

Editable Visual Design combines a vision-language model, an image generator, and a coding agent to produce visual artifacts that remain editable after generation. Its workflow treats generated images as assets inside a structured HTML/CSS design rather than as a final bitmap.

A new paper introduces BCIT to address a central problem in autonomous LLM post-training: whether a previously successful update is still valid after the model and training context have changed.

Google is connecting its Gemini Spark personal agent to Google Photos, allowing eligible subscribers to manage albums, edit images and trigger related tasks with natural-language prompts. The rollout will initially be limited to U.S. users working in English.

OOMOL Lab has open-sourced Open Flow, a workflow automation platform designed for AI agents. It combines a visual Workbench, command-line interface, and self-hosted runtime so agents can help create, orchestrate, and run workflows.

DisCo distills operational knowledge from machine-learning repositories into compact, reusable, and verified skills for research agents. The paper introduces a library with more than 5,000 skills derived from 1,000 repositories.

HarnessDev moves agent evaluation beyond task answers to the runnable infrastructure that shapes those answers. Its results show promising but uneven progress: model-built harnesses still trail mature human systems in coding, search, and research, while performing competitively in writing and machine-learning experiments.

Meta is offering a contributor pricing tier for Muse Spark, its model for coding and other agentic tasks. Users and companies that share prompts and model outputs can receive discounts averaging about 95%.

UI-Venus-2 targets three persistent obstacles in real-world GUI automation: narrow environment coverage, fragile task construction, and unreliable outcome verification. Its unified reasoning–action loop is designed to operate across mobile, web, and desktop interfaces while adding safeguards for consequential actions.

Uber has outlined how it scaled AI agents across software development without allowing spending to grow at the same rate. The approach focuses on routing, context, tools, and measurable workloads.

CAST turns sparse task outcomes into action-level critique supervision for long-horizon tool-calling agents. The approach trains a critique model and uses its judgments to improve the policy model’s reliability across repeated, stateful interactions.

End-to-end data-science automation depends on more than the underlying language model. DS-Lighting turns the often-hidden task, execution, artifact, and evaluation machinery into a reusable harness with four explicit layers.

LLM agents can improve themselves by changing prompts, tools, and execution harnesses, but successful changes may become impossible to undo in later states. EvoUndo studies how precise state grounding and expressive recovery languages affect verified reversibility.

StarHarness treats an agent’s surrounding harness as an optimization target rather than assuming every failure requires a stronger model. By searching prompts, tools, skills, and execution logic for a specific environment, it reports substantial gains across enterprise benchmarks.

Generating code is becoming dramatically faster, but enterprise delivery remains constrained by context gaps, compliance checks and collaboration handoffs. Xiaohongshu’s Muse offers an Agentic approach to connecting ideation, design and production engineering.

Meta explored an AI-native reorganization that would give agents responsibility for large parts of routine work. Instead of removing work, the rollout appears to have increased code noise, incidents, and the burden on engineers who remained.

Cisco is moving its internal AI from a chat interface to personal agents that can retain context, call tools across enterprise systems, and keep tasks moving under supervision. MyAgent is less about the size of the rollout than about making the agent a new entry point for corporate work.

DART-SD models multi-turn tool use as a diamond-shaped state topology rather than a single linear trajectory. It identifies where a rollout goes wrong and distills only the recovery steps that follow.