Separate the session.
Control the effects.

A rejected draft can still leave changes in the workspace.

Fusion can use one model, escalate through a quality-checked cascade, or run draft–critique–revision. Administrators can restrict its model pool separately from general model availability. Current documentation says the main conversation stays on the same model where possible for caching, while assisting reviewers receive only needed context.

Evidence & limits. The documentation explicitly says discarding a draft does not automatically revert its file edits. Fusion is separate from subagents and does not launch them. GitHub’s benchmark and cost comparisons are vendor-reported; they do not establish universal savings or independently reproduced quality.

One host owns the session while several clients can observe and interact with it.

AHP supplies a state snapshot plus ordered actions. Reconnecting clients replay missed actions or obtain a fresh snapshot, then reconcile pending actions. The host arbitrates conflicts; clients negotiate optional capabilities such as terminals and automations. The public protocol repository is MIT-licensed.

Evidence & limits. The architecture was introduced August 26; October 28 is the 1.0 milestone, not its invention. A running host is still required for remote execution. Protocol documentation is not an independent reliability test, and AHP does not standardize each harness’s internal context management.

Inference location and operational control are separate boundaries.

The release adds managed Windows support for MAI Code 1.1 Flash; Copilot CLI also supports manually configured eligible models on Windows, macOS and Linux. Ordinary BYOK discovery does not automatically make a model eligible for routing. GitHub explicitly says routing decisions remain in the cloud, so these sessions are not offline.

Evidence & limits. This is experimental and restricted to supported models and configurations. Intelligent local routing in the Copilot app currently supports managed Windows ML models; directly selecting other local models is a different capability. No air-gap guarantee or complete data-control assurance is established.

Separate proposed actions from the jobs that execute them.

Markdown workflows compile into Actions YAML and reuse runner groups and policy constraints. The agent runs read-only by default behind a sandbox and network firewall. Structured safe outputs are executed by separate permission-controlled jobs. The GA release adds runner support, budget forecasting and isolation options.

Evidence & limits. Current threat-detection docs describe configurable enforcement: default warn mode annotates reviewable outputs, converts branch pushes to review PRs and aborts non-reviewable outputs; strict mode blocks safe outputs on detected threats or detector failures. Detection can be disabled. Human review depends on configuration and policy gates. These are documented controls, not proof of immunity to prompt injection.

Intake, risk tiers, review routing and decision evidence belong in one lifecycle.

Trustible describes a workflow from intake through risk assessment, appropriate review, recorded decision and a centralized inventory. Reviewers can override a risk score with a documented rationale. Each use case retains a lifecycle record rather than ending at committee approval.

Evidence & limits. This is a vendor-reported implementation, not a completed outcomes study. Monitoring signals, metric owners and thresholds for renewed review are still being explored. The monitoring module connects evaluation signals to governance decisions; it does not sit in the runtime path or change model outputs. No measured approval-time or safety gains are reported. The agreement is not an MGB product endorsement.

Formal publication is new this day; the announcement was October 26.

The final rule removes source-attribute support, access and modification requirements and predictive decision-support intervention risk management from §170.315(b)(11)(iv)–(vi) of the ONC Health IT Certification Program. Evidence-based decision-support feedback tools remain. The rule generally takes effect December 28, 2026; specified other amendments take effect January 1, 2027.

Evidence & limits. This is a final rule, not a proposal or an already effective change. Its scope is a particular certification program, not all clinical-AI oversight. Claimed burden reductions and concerns about lost transparency are not measured deployment outcomes. The accessible Federal Register HTML was checked; its linked official PDF was not retrieved.