Three developments put constraints around autonomous AI in sharper focus. Anthropic split its managed-agent harness from durable sessions and sandboxes; OpenAI paused a UK compute plan when energy and regulation changed the economics; SandMLE shrank training environments so reinforcement learning could run against full machine-learning workflows.
1. Anthropic separates agent sessions, harnesses, and sandboxes
Anthropic described Managed Agents as three independent interfaces: an append-only session log, a replaceable harness, and one or more execution sandboxes. Moving the harness outside the container lets a failed sandbox be reprovisioned while the durable event stream preserves progress.
The security gain comes from credential placement. Anthropic keeps OAuth tokens in an external vault and proxies tool calls, so generated code cannot read the secret from its environment. The same separation reduced company-reported p50 time to first token by roughly 60% and p95 by more than 90% through lazy sandbox provisioning.
Sources: Anthropic on scaling managed agents
2. Energy prices and regulation pause Stargate UK
OpenAI paused Stargate UK, a project announced with Nvidia and Nscale that initially targeted up to 8,000 GPUs and could have scaled to 31,000. The company cited regulation and energy cost as conditions for resuming long-term infrastructure investment.
The pause exposes the option value inside infrastructure announcements. GPU targets carry little weight until power, grid access, and policy are contracted. In Britain, high industrial electricity prices and an unsettled copyright regime changed the investment case before the first tranche became durable capacity.
Sources: CNBC on the paused UK Stargate project · OpenAI's Stargate announcement
3. SandMLE cuts machine-learning agent rollouts with micro-datasets
A preprint introduced SandMLE, which generates machine-learning engineering environments with only 50 to 200 training samples per task. The authors report more than a 13-fold reduction in execution time, enabling on-policy reinforcement learning across Qwen3 models from 8B to 30B-A3B.
Micro-datasets make full pipeline feedback affordable, but they also create the transfer question. SandMLE reports relative medal-rate gains of 20.3% to 66.9% on MLE-bench-lite and up to 32.4% higher HumanRank on MLE-Dojo. Those cross-environment tests carry more weight than the synthetic training score alone.
Sources: SandMLE preprint, version 1