AI & Automation Briefing - October 16, 2026
Four Gemini Agents Complete Minecraft Autonomously in 27 Hours
Researchers from Google DeepMind placed four Gemini 2.0 Flash Experimental (Argon) agents in a Minecraft survival world with a single objective: defeat the Ender Dragon. The agents controlled individual players through the open-source Mineflayer library, coordinating via in-game chat and shared memory. No roles were pre-assigned and no game logic was written for them. The agents planned, wrote, and debugged all 2,783 programs themselves, completing the run in 27 hours and 14 minutes with zero human intervention, despite dying 605 times. Researchers could monitor progress on a live dashboard but had no reason to step in. The paper claims this is the first fully autonomous, end-to-end Minecraft completion by AI agents, achieved without pre-scouted seeds or human-built tooling. The coordination pattern, where agents divided tasks, shared learned information, and recovered each other's items across thousands of dependent steps, maps directly to multi-agent orchestration challenges in enterprise workflow automation.
AI Agents Autonomously Post-Trained a Base Model to Match Its Official Instruct Version
Researchers built Athanor, an open system that lets frontier AI agents run the full post-training pipeline on a raw base model without any human guidance. In a 104-hour run, four Claude Opus 5.5 agents working across separate eight-GPU nodes took Qwen3-4B-Base from a score of 9.2 to 48.2 on a held-out evaluation suite, compared to 47.9 for the officially released Qwen3-4B-Instruct. The agents handled everything independently: building training datasets, writing and modifying training algorithms, tuning hyperparameters, merging each other's model weights, and running rounds of self-distillation. No external teacher models were used. The researchers describe this as the first demonstrated case of a frontier agent team autonomously post-training a raw base model into a cross-domain generalist that matches or beats the official instruct release. For teams building automation workflows, this signals that the underlying model layer is gaining the ability to improve itself, which will raise the capability ceiling of the tools built on top of it.
Anthropic Gives External Safety Researchers Employee-Level Access to Audit Models and Risk Practices
Anthropic has granted six researchers from METR, an AI safety nonprofit, access equivalent to internal employees, including company laptops, office space, and the same internal tools used by Anthropic's own risk assessment teams. METR can independently publish assessments of Anthropic's models, incidents, and safety practices without Anthropic's approval. Anthropic retains only a narrow right to require redactions covering customer data, attorney-client privileged material, and model architecture details. METR can publicly note when redactions occur and whether it considers them justified. This arrangement places METR at the top of a three-tier evaluator system Anthropic is building, with lower tiers receiving less access and more publication restrictions. The move follows a September commitment by Anthropic CEO Dario Amodei to embed outside evaluators inside AI labs. OpenAI has made a similar pledge, and both companies along with Google, Meta, Nvidia, and xAI have signed a White House accord committing to independent review. For teams running Claude in production, this structure is directly relevant to compliance conversations with clients and to building a credible position on responsible AI deployment.