ROOT FORWARD / NOVEMBER 20, 2026
Where AI is going—and what you can do with it.

A new model, a selective toolbox, a training tournament, a review proposal, and a local AI desktop all raise a practical question: who can check the result? Today’s five developments offer ways to improve the work while leaving room to question its first answer.
The value of a second opinion is the possibility of a different answer.
Root Note
01 / Models and capabilities
Gemini 4 Neon adds another choice for sustained professional work
Google introduced Gemini 4 Neon on November 18 for complex coding, analysis, and repeated professional workflows. The release brings a one-million-token context window and adjustable thinking levels. Google describes it as a balance between Carbon’s heavier reasoning and Hydrogen’s everyday execution, with support for work involving documents, spreadsheets, charts, audio, and video.
Availability needs a specific destination attached to it. Google lists the paid Gemini API, Gemini for business, and supported Pro or Ultra access in Antigravity and the Gemini app. The launch includes benchmark results and elaborate demonstrations, but neither establishes the quality or total cost of your own completed job. Context capacity also says how much a model can receive, rather than how reliably it will use every detail.
Root Access Take · The useful test is whether a model can stay oriented while the work changes. For a service team, that might mean revising a deployment plan after a dependency moves or a customer changes a constraint. Score the final artifact and its consistency with the revision, including the review work needed to accept it.
Try next · Use an approved, sanitized document bundle and ask for a short implementation plan. Introduce one mid-task correction. Check whether the final plan updates every affected dependency, identifies missing information, and preserves the agreed scope. Compare reasoning settings before choosing a default.
Sources · Google launch, capabilities, and access, November 18
02 / Tools and integrations
Gemini Tool Retrieval loads the instructions an agent actually needs
Google’s November 19 Tool Retrieval preview addresses a familiar agent problem: an enormous toolbox can consume context before useful work begins. Deferred tools retain their names and descriptions, while their parameter schemas load when requested. Developers add tool_search and mark selected tools with defer_loading. Frequently used tools can remain loaded from the start.
The documentation confines this feature to the Interactions API on v1beta; generateContent does not support it. Google lists Gemini 3.6 and later models, including Gemini 4, and support for remote MCP servers. There is no separate retrieval fee, but model tokens still cost money, including loaded schemas. The savings depend on how much of the library a particular conversation actually uses.
Root Access Take · Tool discovery belongs beside good descriptions and narrow permissions. Loading a schema later does not decide whether a caller may execute the action. Keep authorization in the application and expose only appropriate tools. A smaller prompt is useful; an agent that confidently selects the wrong customer’s operation is still a very expensive filing error.
Try next · Compare eager and deferred loading on a small library of read-only test tools. Include two similarly named operations and a request that should find no suitable tool. Inspect the retrieval steps, selected schema, final call, latency, and total token use before expanding the experiment.
Sources · Google preview and billing explanation, November 19 · API version, supported interface, and setup
03 / The MSP workbench
Cisco’s AI tournament offers a training pattern worth borrowing
AWS’s November 19 account of Cisco’s AI League describes 1,200 engineers entering a 2026 agent-building tournament. Participants received preparatory training, built against automated feedback, and progressed through competition rounds. The program combines a workshop, repeated submissions, and a live finale. That is an account of a particular learning program, rather than independent evidence that tournament scores predict production performance.
The winning design supplies the more useful engineering lesson. It used one agent with a concise dispatch approach and placed route planning in a deterministic Lambda tool. Coordination overhead mattered under the game’s scoring rules. The lesson is to examine the workload and its incentives before assuming a more elaborate agent architecture will perform better.
Root Access Take · A small MSP does not need a stadium-sized leaderboard to make this useful. Give technicians protected time, a bounded problem, repeatable feedback, and a chance to explain the tradeoffs. Reward a clear handover as well as the result. Otherwise the team may graduate with a winning prompt that only its author understands.
Try next · Run a short internal exercise using fictional tickets and approved tools. Compare a simple rules-based baseline with an agent-assisted approach. Score routing quality, ambiguous-case escalation, usage cost, and whether another technician can reproduce the result. Finish with a walkthrough of failures and an editable runbook.
Sources · AWS/Cisco case study and winning architecture, November 19
04 / Review and accountability
AI-assisted review needs a route for someone to disagree
In a November 19 UC San Diego interview, researcher Zhe Zhang describes a proposed journal-managed AI review process: assess a submission, let its authors challenge mistakes, then give human reviewers the paper, assessment, and response. The underlying Management Science paper was published online August 7. This week’s explanation makes an older research proposal useful for thinking about everyday review workflows.
The researchers examine a tradeoff: a shared AI assessment may save effort, but can also anchor people to the same initial opinion. Their mathematical model explores conditions for better decisions; it does not demonstrate universal gains in live organizations. Zhang says experimentation is beginning. For broader governance, NIST AI RMF 1.0 remains voluntary and under revision; ISO/IEC 23894:2023 provides risk guidance, while ISO/IEC 42001:2023 specifies AI management-system requirements.
Root Access Take · A human approval button is weak oversight if the reviewer sees only a confident conclusion. Make the evidence and the challenge route visible. The person affected should be able to correct a mistaken premise, and the reviewer should be able to record an independent reason. This proposal supplies no certification.
Try next · Pilot the pattern on a low-stakes internal document. Ask one reviewer to form an initial view before seeing the AI assessment. Record disagreements, corrections, and the final rationale. Check confidentiality and tool permissions before sharing any unpublished or customer material with a model.
Sources · UC San Diego interview, November 19 · Management Science paper and publication date · NIST AI RMF status · ISO/IEC 23894:2023 · ISO/IEC 42001:2023
05 / Local AI and hardware
ASUS puts an always-on local agent into a compact desktop pitch
ASUS’s November 19 ProArt GR1X announcement describes a compact Windows 11 desktop built around NVIDIA RTX Spark, with 64GB or 128GB of unified memory. The company positions it as an always-on host for local agents and creative workloads. Its announcement lists native CUDA support on Windows, dual-fan cooling, and tools such as Hermes Agent and ComfyUI in the surrounding ecosystem.
This is a manufacturer availability announcement, with region-dependent configurations and sales availability. The advertised peak AI performance is a low-precision specification, not measured end-to-end performance for your model. Likewise, a local computer can host a workflow that still calls cloud services. Privacy depends on the selected model, integrations, telemetry, and actual data paths.
Root Access Take · A small local AI box becomes infrastructure once it runs unattended. Someone owns updates, storage, recovery, and the agent’s permissions. Memory capacity helps determine which models might fit, but it does not establish useful response speed or application compatibility. Budget for the whole service rather than treating the desktop as a completed deployment.
Try next · Before purchasing, verify the exact regional configuration, model format, software architecture support, and measured workload requirements. For a pilot, start with public files and a read-only agent. Observe outbound connections and sustained performance, then document restart behavior, access boundaries, and how another person stops the job.
Sources · ASUS availability announcement and regional caveat
Explore the Root Forward archive, AI Radar, Governance guide and The Root Shelf.