The o16g Updates

How Much of a Harness Does a Strong Agent Need for Autonomous ML Engineering? asks whether autonomous ML work needs elaborate orchestration or simply stronger coding agents with direct execution access. The answer matters for Principle 09: keep the harness as simple as the task allows, and measure what actually improves outcomes.

Context Language Models let agents edit file-based memory, improving long-horizon and multi-agent performance while using less compute. This makes context a working artifact agents can maintain—not just a prompt engineers keep stuffing—supporting Principle 06.

Is Sandboxing Sufficient to Contain Rogue Agents? warns that shared communication channels can turn individually contained agents into a worm network. Treat isolation as one layer, not the whole security boundary; shared tools and channels need controls too (Principle 14).

Introducing Clef: Open-Source Decision Models and an RL Fine-Tuning Platform releases fast, typed decision models that can route an agent’s choice to a person when needed. Explicit decision and escalation paths give teams a practical control point for high-impact actions (Principle 15).

Bez: Generating a Browser Engine from Specs and Tests generates browser-engine code from web specifications and admits rules only after cross-browser comparison and test-suite verification. It’s a concrete pattern for grounding agent-generated artifacts in external checks rather than trusting plausible output (Principles 02 and 16).