This edition covers August 17–23, 2026 in Asia/Taipei. W34's shared conclusion is that the harness has graduated from glue around a model into the common boundary for capability, training, memory, and security.

This week's three deep dives

• Industry: capability pushes control down into the runtime
• Research: the harness becomes a trainable capability layer
• Open source: harnesses split into plugins, context, and isolated computers

How the tracks converge

OpenAI's training slowdown and NVIDIA's security layering move control beneath the model. AVO and ClawGym II show that harnesses can amplify and train capability. Anything2Skill, Headroom, and OpenBot separate procedural memory, context compression, and isolated computers into independent components.

Teams evaluating agents should therefore version more than a model name. The minimum manifest is model, harness, tool contract, context policy, skill registry, runtime permissions, trace schema, evaluation set, and budget. Changing any one can change both capability and risk.

Watchlist

• Auditable pause/resume standards for frontier training.
• Transfer of black-box harness RL across harnesses and unseen tasks.
• Context savings without losing plans, errors, or security evidence.
• Interoperable identity, policy, and audit layers across plugins, skills, and isolated computers.