ROOT FORWARD / NOVEMBER 18, 2026
Where AI is going—and what you can do with it.

A coding model, a CRM plugin, an engineering engagement, an industry initiative, and a robot research project all face the same next question: what happens when someone expects useful work? Today’s five developments offer practical ways to test that transition.
A useful handover lets the next person check the work without calling the builder.
Root Note
01 / MODELS / EVERYDAY CODING
MAI Code 2 separates routine coding from the longest jobs
Microsoft introduced MAI Code 2 and Code 2 Flash on November 17. Both are available in GitHub Copilot. The model-specific announcement describes Code 2 Flash in Microsoft Foundry as a limited-access preview, with other distribution channels planned. That distinction matters when a keynote slide makes an entire model family look equally available.
Code 2 targets everyday enterprise development, including changes across files and user interfaces. Flash targets smaller edits, explanations, and rapid iterations. Microsoft explicitly says these models are not yet tuned for fundamental science or very long agentic jobs. Its favorable cost comparisons come from particular benchmarks and configurations; the notes include promotional pricing assumptions. They give you candidates to evaluate, rather than a forecast for your next invoice.
Root Access Take · Match the model to a bounded job before changing the default for every developer. A quick documentation edit and a repository migration need different acceptance criteria. Include the reviewer’s time: a cheaper generation that creates another round of review can make the whole task cost more.
Try next · Give each model the same small change in a disposable repository. Keep tests and review criteria fixed. Record accepted changes, total usage, elapsed time, and minutes of human correction before deciding which jobs to route where.
Sources · Microsoft model announcement and access details · Ignite model lineup and evaluation context
02 / INTEGRATIONS / BUSINESS APPLICATIONS
Salesforce in Copilot requires confirmation for CRM writebacks
Microsoft and Salesforce announced Copilotforce on November 17. Its Salesforce in Copilot plugin entered public preview with more than 38 prebuilt sales skills. The announced examples include preparing for meetings and turning call context into proposed follow-up activities. Supported CRM writebacks require explicit user confirmation and respect Salesforce validation rules, according to the joint announcement.
The surrounding integrations have different maturity levels. Sales Cloud in Teams is a pilot; an open beta is planned later this year. Querying supported OneLake data from Salesforce Data 360 is generally available, while the reverse Data 360 shortcut into OneLake is in preview. Customer agreements govern regional availability and access. Before a pilot, have the tenant and CRM owners verify eligibility, permissions, and permitted actions for the actual users.
Root Access Take · The useful change is bringing a proposed action closer to the context that explains it. Keep the confirmation specific enough to inspect: which customer, which fields, and what source supports the update. A generic approval button can turn careful review into a reflex.
Try next · In an authorized test environment, use a synthetic account and deliberately ambiguous follow-up notes. Check the proposed record change, cancel it once, and confirm a permitted change once. Inspect the CRM afterward for exactly the intended result.
Sources · Joint announcement, action controls, and availability
03 / MSP PRACTICE / SERVICE DESIGN
AWS expands its embedded-engineering offer around 45-day sprints
AWS said on November 17 that Forward Deployed Engineering is available to commercial and public-sector customers in 15 countries, including the United States, Canada, and the UK. The service pairs AWS specialists and agents with customer engineers. Each 45-day sprint has a set price under its service terms and is billed through the customer’s AWS account; the announcement does not publish one universal price.
AWS describes building in the customer’s account, using an opinionated stack that includes Amazon Bedrock AgentCore, reusable business context, and operational controls. Customer engineers work alongside the visiting team so they can maintain and extend the result. These are the provider’s service description and delivery approach. They do not establish that every proposed workflow can reach production within one sprint.
Root Access Take · For an MSP, the interesting product is the capability left with the customer. Put the handover into the statement of work: a named owner, editable business rules, a repeatable acceptance test, and a recovery procedure someone has practiced. Otherwise the monthly service quietly includes calling the original builder forever.
Try next · Take one proposed AI service and write its exit demonstration before estimating implementation. Ask the receiving team to change a rule, investigate an incorrect result, and restore a known-good version using only the materials included in the handover.
Sources · AWS expansion, engagement model, and service terms context
04 / GOVERNANCE / INDEPENDENT EVIDENCE
The AI Transparency Initiative puts independent verification on the agenda
Microsoft announced the AI Transparency Initiative on November 17 with more than 40 participating organizations. Its proposed common approach asks organizations to observe agent activity, verify actions independently, and control delegated authority across models and providers. A useful detail is evidence provenance: distinguish what a model reports from what a runtime, identity system, or separate service actually observed.
This is an industry collaboration to develop an approach and standards, with future product incorporation described by Microsoft. It is not an enacted requirement or a completed certification standard. Keep the established references straight: NIST AI RMF 1.0 is voluntary guidance and is undergoing revision; ISO/IEC 23894:2023 provides AI risk-management guidance; ISO/IEC 42001:2023 specifies organizational AI management system requirements. Joining an initiative or buying a platform does not certify an organization.
Root Access Take · Start with an action you already need to explain. Can the service receiving a request independently show who acted, under whose authority, and what changed? An eloquent agent summary is useful context, but it should be possible to compare it with the receiving system’s record.
Try next · For one synthetic workflow, collect the agent narrative, runtime events, and destination record separately. Introduce a rejected action and check whether the evidence distinguishes an attempted change from a completed one. Assign an owner to any gap.
Sources · Microsoft initiative announcement and scope · NIST AI RMF · ISO/IEC 23894 · ISO/IEC 42001
05 / ROBOTICS / SIMULATION TO REALITY
PEAR trains robot-use reasoning in simulation, then tests real arms
A November 13 preprint introduces PEAR, a training system for teaching a reasoning model to operate robots through reinforcement learning in simulation. The team starts from Qwen3.8-27B. During this robot post-training it uses no demonstrations or real-robot data. The model observes a scene, reasons, and emits the next arm pose. Public code and model releases make the approach worth examining beyond the demonstration videos.
The researchers report 18 successes in 30 real dual-arm trials after training in LIBERO and two RoboDojo task types, compared with none for the base model. Those trials cover pickup and block alignment; they do not represent a general factory workload. The team also reports slow execution because the model reasons before each move. Its fast, lower-precision decoding examples use a different configuration from the BF16 evaluations. Keep throughput demonstrations and measured task success separate.
Root Access Take · The promising result is transfer from simulated practice to unfamiliar physical hardware. The operational question is how the system fails when an object, camera, or initial arm position changes. A robot still needs a carefully bounded workspace and a way for a person to intervene.
Try next · Begin with the released simulation setup and study failure cases before considering physical experiments. Track success, reasoning delay, and recovery behavior separately. Require a fresh assessment for the particular robot and workspace before moving beyond simulation.
Sources · Research preprint, November 13 · Team’s methods, trial breakdowns, and limitations · PEAR code and setup
Explore the Root Forward archive, AI Radar, Governance guide and The Root Shelf.