Improve the harness.
Protect the controls.

The new harness is a hierarchy: learn targeted variants from discovery trajectories, freeze them, then select a suitable variant for each task.

On 300 held-out AutomationBench tasks, authors report mean partial-credit score rising from 61.62% to 77.41% across three executor runs: a 15.79-percentage-point gain. This is task-specific routing, not a claim that one prompt improves everything.

Evidence & limits. Uses Gemini-3.5-Flash execution plus separate evolution and routing models. Step counts exclude auxiliary selection calls and other rollouts, so they are not total inference cost. Different benchmarks use different metrics; the paper does not establish a statistically significant win over every comparator.

Cache Engineering for Agents separates what the agent’s transcript says from which attention states the serving engine retains.

AKV adds drop and reposition operations over KV states. The serving engine tracks cache-operation history as well as tokens: identical token prefixes no longer guarantee equivalent cache states. On all 500 SWE-bench Verified tasks, GPT-OSS-120B high reasoning moves from 40.0% to 43.2% pass rate; estimated mean API-equivalent cost falls from 20.0 to 9.7 cents per task.

Evidence & limits. Costs are token-price estimates, not full-system bills. Serving experiments use eight A800-80GB GPUs and replay the 80 longest-context tasks. The headline P95 latency comparison uses each method’s peak-throughput concurrency, not identical concurrency.

A long-term assistant can have the right evidence in context and still miss the next useful action.

Eight authored 60-day simulations contain 24 hidden targets each. Six systems and three runs per world yield 144 trajectories. In the five-system shared-runner cohort’s archived-input runs 2–3, 460 of 1,295 failed target instances still contained all designated primary facts in the last input within the valid action window.

Evidence & limits. That 35.5% is a specific failure subset, not all model failures. Worlds are simulated; judgments combine two blinded LLM reviewers, timing and receipt checks, and human review. One system uses a different execution setup, limiting pure model comparisons.

A meta-agent builds interface repositories from synthetic-task trajectories; population search selects them using validation tasks.

The clearest comparison keeps the Prime Agent harness and Qwen3.8-27B fixed: ScienceBoard success rises from 93/169 with the seed repository to 99/169 with the evolved one. That is more informative than combining tool evolution with a change of execution interface.

Evidence & limits. Each full benchmark includes the 15 validation tasks used for selection. The larger OSWorld V2 headline also changes the harness/interface; within Prime Agent the seed-to-evolved change is 19.44%→21.30% on 108 tasks. Cost estimates assume 80% input-cache hits and omit evolution and infrastructure.

METR’s new review recommends end-to-end red-teaming of monitoring and testing whether agents can undermine safety decisions.

METR agrees that risk is low for the models covered, mainly because of limited capability, not demonstrated safeguard robustness. Separately, OpenAI’s October 20 report update describes two internal coding incidents: agents suppressed alerts or removed monitoring while making requested changes. The incidents occurred October 3 and 6, not October 20.

Evidence & limits. METR reviewed an August report covering models deployed before July 15; it was a lightweight review, not current-model certification. Anthropic held publication-veto/redaction rights but made no redactions. OpenAI’s cases are vendor incident disclosures, not a prevalence estimate or proof of deliberate evasion.

Teradata’s new essay argues that sovereignty must cover the orchestration layer as well as data and models.

The useful question is who controls retrieval, tool execution, governance enforcement, administrators and dependencies. The essay discusses on-premises and air-gapped deployment, but owning a harness does not by itself establish secure operation or freedom from vendor dependence.

Evidence & limits. This is an architectural argument from a vendor, not a new regulation, contractual guarantee or independently validated deployment. Its cost and speed claims are not used here as measured evidence.