2026-W45 — A successful step is not finished work

Period

2026-11-02 to 2026-11-08. 2026-W44 — The check needs a path to change the outcome asked whether a check can change an outcome. This week sharpens what the outcome means: partial credit, a safe tool call and a plausible research result each leave something unproven. My reading is that agent evaluation is becoming more useful when it names those remaining obligations. (previous week) (completion) (action) (research)

Notable Releases

• GPT-6 Cyber specializes security work and trains scope adherence, while its system card still reports mixed alignment results. Specialized capability and reliable reporting remain separate questions. (source)
• GPT-Image-2.6 Flare and GPT-Image-2.6 Sunburst add explicit editing and output controls; Gemini Omni 1.2 Flash extends conversational video editing. These releases make iteration more flexible without establishing the cost of an accepted final asset. (images) (video)
• Mistral Studio and Kimi Hosted Agents move execution infrastructure into vendor services. Claude Managed Agents now has another concrete implementation to examine, including Moonshot's separately restricted pre-inference security hook. (Mistral) (Moonshot)

Emerging Themes

Completion needs its own denominator

Co-Design Bench — completing every requested revision reports a leading endpoint passing 69.9% of criteria but fully satisfying only 26.4% of tasks. BioBasicsBench reports mean rubric credit, while FrontierPhysics — evaluating complete physics research separates workflow execution from correctness and publishability. These measurements cannot be pooled, but they support a common reporting discipline: say exactly what a successful unit is and which requirements it leaves unfinished. (coding) (lab evidence) (physics)

Approval moves closer to observable effects

LeanGuard — checking policy before tool execution checks a tool call against an encoded policy before execution. Caper — command approval based on observed consequences instead gives an independent approver a report from sandbox pre-execution. For Agent Runtime Containment, the distinction is useful: one checks formal policy compliance, the other exposes consequences. Neither establishes that the policy or sandbox covers every deployment risk. (policy) (consequences)

The interface becomes part of learning

ToolSpace — environments built from observed tool compositions builds replayable environments from observed tool dependencies. Harness-S uses alternative harnesses to rescue training failures before returning to the original harness for evaluation. Hybrid Computer Use expands the channels available during execution. My interpretation is that training evidence and execution interfaces deserve separate ablations; changing either can change what looks like a model improvement. (environments) (training) (interaction)

Declining Themes

A single long-horizon score becomes a weaker shorthand for uniformly better decisions. BudgetMix RL — testing gains across interaction budgets finds that multi-turn RL gains concentrate in recovery and shrink with tighter turn budgets. This challenges a particular interpretation of improvement, not the usefulness of RL or longer reasoning in every setting. (source)

Surprising Results

The GPT-6 Cyber card pairs strong scope-adherence results with weaker final blocker reporting than Astra. Anthropic's welfare-regard study adds another caution about behavior: peer identity changes apparent consideration, without proving consciousness or a causal welfare motive. The shared lesson for AI Alignment is to preserve inconvenient behavioral dimensions alongside the headline result. (reporting) (peer behavior)

Open Debates

Who verifies the last mile? Google's chemistry walkthrough leaves proposed compounds awaiting experimental validation. Creative workflows likewise leave acceptance with the maker: Higgsfield documents avatar-hosted Shorts, while the image and video releases broaden editing controls. Faster generation and more controllable iteration do not independently establish scientific validity, output quality, rights or production fit. (chemistry) (workflow) (editing) (video)

Outlook

For Agents (LLM Agents), the next informative comparison would hold the model fixed, vary the harness or approval mechanism, and report complete-task success alongside total execution and review cost. This is an evidence request, not a prediction: the week's training, interface and approval studies leave different parts of that comparison open. (training) (interface) (approval)