ROOT FORWARD / NOVEMBER 11, 2026
Where AI is going—and what you can do with it.
A model gets a deeper review pass, an AWS API hands back the search results, and Quick packages a repeatable service-desk task. Stanford asks who should help shape AI deployments, while NVIDIA measures how many demonstrations a robot needs to learn a precise move.
The receipt that matters is the one someone else can use to check the work.
Root Note
01 / MODELS / REASONING
Gemini 4 Carbon puts a price on the extra pass
Google made Gemini 4 Carbon available on November 10 through the paid Gemini API and Gemini for business, with a rollout to Google AI Ultra subscribers across selected Google products. Google positions Carbon for longer, more demanding coding and knowledge work. It also says Carbon uses more tokens than Gemini 4 Argon to explore a problem and verify its work.
The public API’s introductory price is $2 per million input tokens and $10 per million output tokens. Google says those rates rise to $4 and $20 after November 30. That makes a trial in November useful for measuring quality and consumption, provided the business case uses the December rates. Google reports gains on several workflow benchmarks. Your own acceptance tests still determine which model fits the job.
Root Access Take · Give the expensive pass a job description. A migration that spans several repositories may warrant a second model to check dependencies and assumptions. A routine documentation edit probably needs a good diff and a competent reviewer. Price the work you expect people to accept, including retries and review.
Try next · Replay a sanitized multi-file refactor with Argon and Carbon under the same tool permissions and test suite. Compare passing checks, tokens, elapsed time and review minutes. Calculate cost per accepted change using the announced post-introductory rate.
02 / INTEGRATIONS / RETRIEVAL
AgentCore exposes web search as a direct API
On November 10 AWS made AgentCore’s InvokeWebSearch API generally available. Applications can call it through the AWS SDK or CLI using IAM permissions. The same search index was already available as an MCP tool through AgentCore Gateway. The new path lets a program or pipeline retrieve results directly without provisioning a gateway.
The response is a ranked list of extracts, with titles, URLs and publication times when available. Request filters can restrict domains and dates. AWS lists six supported regions, including Northern Virginia and Ireland. The direct API uses IAM and per-request domain filters. A gateway can enforce target-level domain lists and supports a different authorization model. This search endpoint returns evidence for your application to inspect, while the separately available Web Answer tool synthesizes a response.
Root Access Take · Keep a search result attached to the claim it supports. A recent extract can still describe an old feature, and a high ranking tells you little about whether the publisher owns the information. Keep the original URL and check the source itself before it goes into a report.
Try next · Try five questions about recent release dates against a short list of official domains. Log the result links, missing dates and source disagreements, then count how many claims survive a manual check.
Sources · AWS launch announcement · AgentCore developer guide
03 / MSP PRACTICE / REPEATABLE WORK
AWS Quick shows how to hand a workflow to a team
AWS published a walkthrough of skills and agents in Quick on November 10. A skill packages instructions for a repeatable job in a folder with a SKILL.md file and optional references or scripts. The agent can load the instructions when a request matches the skill or the user names it. Quick on desktop already supports this format. The new post explains how to build, test and share these workflows.
For an MSP, a candidate is the weekly service review. One skill could lay out which ticket fields to read, how to group unfinished work, and how to mark missing evidence. Publishing the skill and sharing it with colleagues are separate steps. The AWS guide also documents administrator controls and connector permissions. A shared recipe still depends on each colleague’s access to the inputs.
Root Access Take · Start with a workflow where a wrong answer is easy to spot and the first output stays in draft. A portable checklist is worth something even if the tool changes. Include the point at which the agent should stop and ask a technician to interpret an ambiguous ticket.
Try next · Give two technicians the same sanitized ticket export and ask them to run the review with the skill. Compare required fields, exception handling and the amount of cleanup before anyone can use the result.
Sources · AWS implementation walkthrough · Quick desktop skills and permissions
04 / GOVERNANCE / ADOPTION
Stanford’s AI100 panel asks who gets a say
The One Hundred Year Study on Artificial Intelligence, administered by Stanford HAI and overseen by an international standing committee, released its third panel report on November 10. AI Goes Mainstream examines the period since the 2021 report. It covers the scaling of generative AI and its effects on areas including work, education and public trust. The panel calls for more input from people affected by AI deployment.
That question scales down to a help desk. A new routing assistant may make a queue look tidier while pushing the hard cases onto the same technician. Ticket averages can miss that redistribution. The AI100 report is a research and policy analysis. Teams looking for operational practices can turn to NIST’s voluntary AI Risk Management Framework, ISO/IEC 23894:2023 for risk-management guidance and ISO/IEC 42001:2023 for AI management system requirements.
Root Access Take · Put the people who get the exceptions in the room before the pilot goes live. Ask what would make the workflow harder to challenge or correct. That conversation is part of designing a service that a customer can actually use.
Try next · For a proposed pilot, pick one review date, one named owner and one way for affected staff or customers to challenge a result. At the review, check where work moved as well as how much time the pilot saved.
Sources · Stanford HAI release · AI100 report scope · NIST AI RMF · ISO/IEC 42001 · ISO/IEC 23894
05 / ROBOTICS / MANIPULATION
GR00T 2 shows what five demonstrations can buy
NVIDIA presented GR00T 2 at the Conference on Robot Learning on November 9. The world action model uses recent video, instructions and robot state to predict future observations and a short sequence of actions. NVIDIA trained it on human video, handheld gripper demonstrations and robot teleoperation data. The aim is to reuse skills across different robot bodies and adapt them to new tasks with relatively little target-robot data.
In NVIDIA’s 50-task study on a two-handed robot, the base model succeeded on 40 percent of trials and completed 40 of the 50 tasks at least once. Those are different denominators. After fine-tuning with one demonstration per task, average trial success rose to 71 percent. Five demonstrations brought that figure to 95 percent in this setup. The research also includes separate assembly and cable-insertion demonstrations on another robot.
Root Access Take · Ask what the five demonstrations cover. A connector that starts neatly on a mat is a different job from a connector hanging in a crowded rack. A useful trial names the fixture, the variation and the recovery procedure as carefully as it names the model.
Try next · For a bench test of a manipulation task, define a clean run, an offset part and a dropped part. Score completion and safe recovery separately, and record which retries needed a person.
Sources · NVIDIA CoRL announcement · NVIDIA research and evaluation details
Explore the Root Forward archive, AI Radar, Governance guide and The Root Shelf.