ROOT FORWARD / NOVEMBER 19, 2026
Where AI is going—and what you can do with it.

New models, recurring tasks, cleaner datasets, agent evaluations, and editable worlds all offer something worth trying. Today’s question is what to put on the test calendar—and what evidence to keep when the experiment finishes.
A recurring job needs an owner, a test, and a result someone can check.
Root Note
01 / Models and capabilities
GPT-6.2 Astra and Sol make model replacement a workflow decision
OpenAI released GPT-6.2 Astra and Sol on November 17, positioning Astra for harder professional projects and Sol for everyday work. Both are available through the API as gpt-6.2-astra and gpt-6.2-sol, alongside supported ChatGPT Work and Codex access. OpenAI reports improvements across coding, computer use, and business documents; those are provider evaluations, not a guarantee about your particular workflow.
One migration detail deserves its own change ticket: the API changelog says both models support Responses and Chat Completions, but tool calling requires the Responses API. A successful plain-text request therefore does not prove an existing agent integration is ready. Model choice, orchestration, and output checks need to be tested together.
Root Access Take · Treat a model upgrade like replacing a component in a service. The useful comparison is accepted work after review, including retries and handover effort. A quicker first answer can still produce a slower completed job. Keep the previous configuration available until the new one clears your own acceptance checks.
Try next · Replay a small set of representative, sanitized tasks against the current and proposed configurations. Include a document, a code change, and a tool action where relevant. Record completion quality, reviewer time, latency, and total usage; separately confirm that the intended API path supports every required tool.
Sources · OpenAI release and availability, November 17 · API changelog and tool-calling requirement
02 / Tasks and integrations
Gemini Tasks expands access, with a consumer-account boundary
Google’s November 17 Gemini release notes expand free access to Tasks in the United States, India, Canada, Australia, Japan, and Korea. Tasks can pursue multi-step projects in the background or run recurring routines, using permitted context and connected services. The practical change is availability: more people can try ongoing work rather than repeatedly rebuilding the same request.
Eligibility matters. Google’s current help page requires a personal Google Account, age 18 or over, and Gemini History enabled; work and school accounts cannot use Tasks. Browser execution also has two different operating conditions. Local Chrome work needs the computer awake and Chrome running. Remote browsing can continue after the device closes, although sign-in can interrupt it. Users confirm browsing tasks before they begin.
Root Access Take · An MSP can learn from a personal-account pilot without treating it as an enterprise deployment. Start with public information and a bounded deliverable. A recurring task needs an owner, a stop condition, and a clear definition of a missed or unusable result. A calendar entry is not an operations team.
Try next · Try a weekly comparison of public vendor documentation using an eligible account. Specify the sources, expected output, completion time, and what to do if a page is unavailable. Check connected-app permissions and confirm whether the task depends on an awake local browser.
Sources · Gemini release notes, November 17 · Tasks eligibility, setup, and browser conditions
03 / The MSP workbench
SageMaker adds data preparation; agreement still needs a human check
AWS introduced SageMaker AI data preparation on November 17, combining automatic labeling, synthetic examples, quality checks, transformations, and human review. Teams can use Python SDK APIs or an AI-ML agent skill in supported development environments. Generated specifications and code remain inspectable and editable, which makes this relevant to service teams building repeatable evaluation datasets.
The service is generally available in Northern Virginia and Oregon. It uses a customer-controlled IAM execution role and stores source data, generated data, and labels in the customer’s S3 buckets. For supported fixed-answer labeling tasks, agreement among several models produces a confidence score. Rows below a chosen threshold can enter human review in SageMaker Studio. That score measures agreement; correlated mistakes remain possible.
Root Access Take · Dataset preparation can become a useful service deliverable if the customer can inspect its origin and limitations. Keep generated examples distinguishable from observed cases. Reserve human-reviewed cases that were not used to tune the pipeline, or your evaluation may reward the same assumptions that created the answers.
Try next · Choose one narrow support classification task with approved, sanitized examples. Have a domain expert label a small reference set, then compare machine labels and disputed cases against it. Review IAM permissions, regional availability, generated code, and expected charges before scaling the workflow.
Sources · AWS launch, workflow, regions, and data controls, November 17
04 / Evaluation and accountability
Copilot Evals gives agent customizations a repeatable check
GitHub’s November 17 public preview of Copilot Evals adds a way to check skills, custom agents, and MCP-based workflows against real tasks. Users can request an evaluation in natural language or with /eval, inspect failed checks, and compare behavior before and after a customization. GitHub says each evaluation run requires approval and uses the local Vally evaluation engine.
Access spans Copilot Individual, Business, and Enterprise plans through the documented VS Code extension, CLI plugin, or app marketplace path. This is a preview tool for gathering evidence. NIST’s voluntary AI RMF 1.0, released in 2023 and now being revised, provides broader risk-management context. ISO/IEC 23894:2023 is AI risk guidance; ISO/IEC 42001:2023 specifies organizational AI management-system requirements. Passing an agent evaluation does not confer certification.
Root Access Take · A skill deserves a test that can disagree with its author. Save the task, expected evidence, environment, and failure criteria alongside the customization. Run it again when the model, instructions, or connected tools change. Otherwise the green check belongs to a configuration you may no longer be running.
Try next · Pick one internal agent customization and define three cases: ordinary completion, missing context, and an action outside its allowed scope. Review the proposed run before approving it. Inspect the resulting artifacts and action sequence, then document who decides whether a failure blocks release.
Sources · GitHub preview, supported plans, and installation paths · NIST AI RMF status and scope · ISO/IEC 23894:2023 · ISO/IEC 42001:2023
05 / Interactive worlds
Genie 4 reaches eligible Ultra users, with simulation limits intact
Google’s November 18 Project Genie update opens Genie 4 to eligible Google AI Ultra users, expanding the limited-tester access covered in Root Forward’s first edition. The browser experience adds worlds that respond to prompts during exploration, including changes to weather, characters, and actions, with native audio. Google AI Pro users can join a waitlist. This is a meaningful access change, not a repeat of the October announcement.
Keep the model boundaries straight: the same update describes longer sessions and world chaining for Genie 3, a separate model choice. Those details should not be silently transferred to Genie 4. DeepMind’s Genie 4 research description also flags drifting visual details and object permanence, imperfect physical interactions, spatial accuracy, and consistency over longer sessions.
Root Access Take · An interactive scene can help people explore an idea without proving that its underlying physics are dependable. Creative prototyping and validated engineering simulation have different acceptance criteria. For an MSP or trainer, the interesting experiment is whether a controllable scene improves discussion or instruction, and where it introduces misleading expectations.
Try next · If your account is eligible, build one short fictional training environment and change a single condition while exploring it. Note which objects and relationships remain consistent. Label the result as an experimental generated world, and avoid using its behavior as evidence for physical equipment or robotics decisions.
Sources · Google access update and separate Genie 3 features, November 18 · DeepMind research description and limitations · Earlier Root Forward coverage of limited-tester release
Explore the Root Forward archive, AI Radar, Governance guide and The Root Shelf.