xAI uploads codebases 👨‍💻, Prime Intellect verifiers 🧠, Sakana smart bricks 🧱

Combine AI reasoning with deterministic execution (Webinar) (Sponsor)

Most agents re-decide their next move at every step, so every run is non-deterministic end to end: hard to debug, hard to govern and hard to trust. This webinar breaks down a different approach: separating the agent's reasoning from how that reasoning gets executed.

What you'll learn:

How agents and workflows call each other, so AI reasoning and business logic aren't stuck in separate systems
How an agent's one-time plan becomes a fixed, durable Orkes Conductor sub-workflow
How Orkes Conductor recovers a failed agent run from the exact point of failure, instead of restarting from scratch
How to bring LangGraph, OpenAI Agents SDK, Google ADK, or native agents into this same model

Join us July 23 to see it in action.

🚀Headlines & Launches

Today, we are releasing verifiers v1 (3 minute read)

Prime Intellect's verifiers v1 is an overhaul of its environment stack for the modern era of agentic RL and evals. It decomposes environments into a task set, a harness, and a runtime. verifiers v1 enables complex agentic tasks like coding and computer use at scale in any harness.

Smart Cellular Bricks: Towards Collective Intelligence for the Physical World (11 minute read)

Sakana AI's research on collective intelligence has extended into physical hardware with "Smart Cellular Bricks." These bricks employ local communication and neural networks to autonomously classify and reconstruct 3D shapes without centralized control. In experiments, the system achieved high accuracy in shape classification and demonstrated robustness against noise and module failures.

🧠Deep Dives & Analysis

What xAI Grok Build CLI actually sends to xAI (20 minute read)

This post looks at the data that xAI's official Grok Build coding CLI transmits and stores. The CLI transmits the contents of the files it reads to xAI verbatim and unredacted. It uploads whole repositories independent of what the agent reads. The storage destination is a Google Cloud Storage bucket. There's no proof that xAI trains on the data.

Making Fable Cheaper Than Opus (12 minute read)

Cognition replaced Opus 4.8 with Fable 5, and Devin's bill went down. Fable 5 costs twice as much per token as Opus 4.8, but Cognition created an architecture that resulted in Fable costing less while scoring higher. This post looks at how the company achieved this and what that means for pricing agentic work.

👨‍💻Engineering & Research

Workshop: Build AI agents with scoped credentials (Sponsor)

Long-lived agent credentials are a breach waiting to happen. This workshop covers agent identity and access management on AWS, including scoped credentials via AWS STS, automatic rotation with AWS Secrets Manager, and audit trails via AWS CloudTrail. Plus technical demos of Auth0 and CyberArk PAM.

Video Generators as General-Purpose Vision Models (8 minute read)

DeepMind researchers introduced GenCeption, which repurposes a pretrained video generation model as a unified vision system controlled through text instructions.

Testing Agents on Long-Horizon Terminal Work (GitHub Repo)

Long-Horizon Terminal-Bench evaluates whether LLM agents can sustain productive work across hundreds of terminal interactions. Its 46 stateful tasks use hidden verifiers that rebuild and inspect final artifacts rather than trusting an agent's reported progress.

How Microsoft Ships Thousands of Production AI Agents (18 minute read)

Microsoft described the infrastructure and key mechanisms needed to operate AI agents across Foundry and its large-scale Copilot products. Key practices included treating retrieval as a sub-agent, assigning agents distinct identities and workspaces, and using rubric-based evaluations with automated improvement loops.

Mantis Skills: Portable Toolkit for Building Security Review Harnesses (GitHub Repo)

Mantis Skill is a decoupled, sequential, and security-focused set of Skills designed for use with Coding Agents. It is intended to be a flexible foundation that users can adapt, tune, and extend, rather than a rigid set of instructions. Google recommends using AI to iterate on the skills and use internal documentation, coding standards, and build systems to augment the threat model. It also strongly recommends adapting risk calibration to the environment and risk tolerance.

🎁Miscellaneous

Better Call Sol: The Workhorse (16 minute read)

Zvi Mowshowitz argued GPT-5.6 Sol delivers the strongest balance of reasoning quality, speed, and cost among current frontier models, making it the default choice for demanding knowledge work. He highlights its long-horizon execution, effective computer use, and agentic workflows while noting that task-specific model selection still matters for peak performance.

The Most Human Technology Ever Made (8 minute read)

AI is emerging as a transformative tool, empowering individuals to create and innovate, rather than just consume. It reduces the barriers to production, allowing people without technical backgrounds to build practical and creative applications. Unlike consumption-focused technologies, AI nurtures individuality and creativity, fostering a culture where personal projects and ideas can flourish and become meaningful contributions.