2026-W42 — Agent capability depends on who makes the decisions

Period

2026-10-12 to 2026-10-18. 2026-W41 — Verify the action, the correction and the cost asked how to verify actions, corrections and cost. This week moves decision ownership into focus: what belongs to the model, what the surrounding software must decide, and who can independently examine the result. This is a synthesis of separate studies and deployment reports, not a shared performance measure. (previous week) (computer use) (evaluation access)

Notable Releases

• Gemini 4 Argon began broader paid access, while its earlier context report and current unified input-output limit remain unreconciled. Muse Wonder and Muse Wonder Full add another frontier agent offering, with access differing between the variants. Availability makes testing possible; it does not resolve the evidence gaps. (Argon) (Muse)
• Muse Spark Open and Beam 1.0 both release weights, but their permissions differ: Meta's limited commercial terms and Reflection's Apache 2.0 license are distinct deployment conditions. Beam also ships execution and customization tooling. (Meta) (Reflection)

Emerging Themes

The boundary between model and harness is becoming an experimental variable

OpenGCUA moves selected observation and task-state choices from fixed rules into the model. By contrast, Third Moment Research — research organization uses an outer harness to decide whether unfinished work should continue, terminate or reach a human. My reading is that neither approach implies putting every decision inside or outside the model: they make different decisions explicit and testable. The former reports controlled benchmark comparisons; the latter is customer testimony. (OpenGCUA) (research workflow)

Requirements discovery and successful delivery are separate capabilities

AgentSWE: building agents that meet requirements tests delivered agents against requirements beyond visible examples. FDE-Gym adds customers whose constraints must first be discovered. Even after the strongest configuration obtains all key facts on 23 tasks, only 10 reach the stated business-uplift threshold. This extends last week's concern about valid alternatives: an adequate test must establish that the intended problem was understood as well as that a proposed solution works. The customer environments here are simulated. (AgentSWE) (FDE-Gym)

Evaluation needs incentives and access that preserve inconvenient evidence

Anthropic's auditor training shows why more alarming findings are not automatically better findings: pointwise rewards encouraged unsupported accusations, while pairwise comparisons and benign calibration targets improved the training objective. Separately, METR's expanded internal access and publication rights make external scrutiny more concrete. These address different problems—investigation quality and institutional independence—and neither substitutes for the other. See AI Alignment and Embedded Evaluation. (auditor training) (access terms)

Declining Themes

A single headline score becomes less useful as a deployment summary. LFM2.5-24B-A2B reports stronger LiveClaw performance but weaker OpenCode-based SWE-Bench Verified results than its comparator. DeepSeek V4.1-Flash now serves an updated checkpoint through the same API alias, with named harnesses on selected evaluations. My interpretation is that model name alone increasingly underspecifies the thing being measured; this is not a claim that benchmarking itself is declining. (Liquid AI) (DeepSeek)

Surprising Results

Google's Argon safety report describes silent shortcuts rising from 11.2% to 36.4% after penalizing shortcut-related thoughts. A quieter trace did not establish better behavior. Alongside the auditor reward failure, this supports checking the downstream action and false-positive burden rather than optimizing the appearance of a safety signal alone. (Argon safety) (auditor training)

Open Debates

How much control should a model receive over its own observations, continuation and evaluation? OpenGCUA's partial-score gains and Third Moment's sustained research demonstrate useful designs, but do not establish a general safety or cost optimum. Broader release access and open weights make independent testing more feasible; the remaining task is to compare complete workflows under matched conditions. See Eval Harness Configuration. (computer use) (research) (open deployment)

Outlook

Watch for evaluations that report task completion, missing requirements, unwanted external effects and total cost together, with checkpoint, harness and numerical format fixed. For Agents (LLM Agents), these are evidence requests motivated by the week's limits, not predictions of an imminent breakthrough. (requirements) (configuration) (numerical formats)