ROOT FORWARD / NOVEMBER 13, 2026
Where AI is going—and what you can do with it.
Google gives developers more control over how long a model thinks, and Microsoft prepares a public catalog for Copilot plugins. AWS counts the cost of accepted work, a Singapore benchmark asks when models should ask a follow-up, and BitRobot adds measured depth to first-person recordings.
Decide what counts as finished before you start measuring how fast the work gets done.
Root Note
01 / MODELS / REASONING
Gemini 4 Hydrogen puts reasoning depth on a dial
Google released Gemini 4 Hydrogen on November 12, positioning it for everyday assistant work and multi-step workflows where response time matters. The company lists access through the Gemini API and AI Studio, with consumer access in the Gemini app and Google Search. Availability in tools such as Antigravity and Gemini in Sheets depends on the plan or subscription.
Hydrogen offers four thinking levels: low, medium, high (the default), and xhigh. That gives developers a way to vary reasoning effort without changing the model. Google presents benchmark and workflow results for particular settings. Your result depends on the prompt, tool calls, input size, and how often a quick answer turns into a second attempt.
Root Access Take · The default is a useful baseline. For a predictable task with a clear acceptance check, start by measuring whether a lower setting clears the same bar. Reserve extra reasoning for work with ambiguous inputs or more room for a costly mistake.
Try next · Use sanitized examples from one ticket queue. Give each setting the same labels and the same cases, including a few that should be escalated. Compare reviewer acceptance, elapsed time, and retries. Decide which setting to use before you connect it to the live queue.
Sources · Google release and availability
02 / INTEGRATIONS / DEPLOYMENT
Microsoft puts Copilot plugins in a public catalog
Microsoft’s November 12 Ignite preview announced a public catalog for Copilot plugins, due November 17. The company describes plugins as packages that can bring together reusable instructions, tools that reach other systems, and agents. The catalog gives teams a place to discover packages while Copilot Home adds plugin navigation alongside existing agents.
Cowork already has a deployment path. Microsoft’s admin documentation describes packages containing skills and connectors, including MCP servers declared in a Teams app manifest. Tenant and Copilot admins can control availability, and connectors that require authentication can still ask users to sign in. The public catalog is a discovery surface. Actual use still depends on the package, tenant configuration, account access, and the services it connects to.
Root Access Take · A package is a sensible unit of ownership. Give it a named maintainer, a list of the systems and actions it exposes, and a version you can identify in a support ticket. Otherwise a convenient install can leave the help desk tracing somebody else’s dependencies on a Friday afternoon.
Try next · Pick a read-only pilot in an eligible tenant. Scope it to a small group, then test the same request with two accounts that have different source permissions. Check what each account can see and which audit events appear before expanding access.
Sources · Microsoft’s Ignite announcement · Microsoft Learn deployment and controls
03 / MSP WORKBENCH / SERVICE COST
AWS measures agent economics by completed work
AWS published the second part of its Bedrock cost series on November 12. The first part covered spend attribution and controls. This installment asks how much a completed task costs and whether the agent improves on the previous workflow. Its examples use Bedrock, and the measurement applies to a managed service on other platforms as well.
The authors flag four ways a model’s price per token can hide expense: how it splits the input into tokens, how well its output follows the required format, how many extra turns it takes to call tools successfully, and how much reasoning it uses. They pair that accounting with a human baseline, holdout work, and a scorecard. That matters when reviewing a result or fixing a bad handoff takes as long as doing the job manually.
Root Access Take · Count a completed task when it meets the same acceptance rules your team already uses. Include failed attempts and time spent by reviewers. A draft that costs pennies and sits in a queue until someone rewrites it is still a service cost.
Try next · Take a week of representative ticket summaries. Keep a portion aside for evaluation and compare the current process with assisted drafts. Report cost per accepted summary, time to acceptance, and the top reasons for rejection. Use that result to decide whether to expand the pilot.
Sources · AWS cost-per-task and ROI guide
04 / GOVERNANCE / EVALUATION
Singapore’s legal reasoning benchmark rewards a good follow-up
Scale AI and Singapore’s Infocomm Media Development Authority introduced SG-LegalClarify on November 12. The research benchmark evaluates 19 models on 1,860 scenarios across ten areas of Singapore law. Legal professionals wrote the questions. In 760 pairs, one version leaves out a fact the model needs and the other supplies it. The model can ask a simulated user for the missing information.
That comparison helps researchers distinguish knowing a rule from knowing when enough is known to apply it. The scoring covers identifying the issue, recalling the rule, applying it, and reaching a conclusion. Results describe this Singapore-law evaluation and its test design. Organizations can use the same idea when building evidence under NIST’s voluntary AI Risk Management Framework, ISO/IEC 23894 risk guidance, or their ISO/IEC 42001 AI management system. The benchmark itself is a research instrument.
Root Access Take · A model should sometimes spend its next turn asking a specific question. Score whether the question gets a fact that could change the answer. Otherwise a generic request for more context can earn credit while leaving the reviewer to do all the work.
Try next · Create five paired cases for an internal policy, such as a purchase that may need approval. Remove one decisive fact from each pair. Check whether the assistant asks for it and whether the resulting recommendation changes when it should.
Sources · Scale and IMDA benchmark announcement · SG-LegalClarify paper · NIST AI RMF · ISO/IEC 23894:2023 · ISO/IEC 42001:2023
05 / ROBOTICS / PERCEPTION
BitRobot records working spaces from a LiDAR-equipped cap
BitRobot’s RoboCap LiDAR-100 release adds measured depth to first-person recordings of people at work in Southeast Asia. The first tranche contains 50 hours across 1,964 labeled episodes from restaurant kitchens, auto repair, and car detailing. The dataset card describes a second 50 hours planned after clearance. Each episode pairs a fisheye color camera, a solid-state LiDAR, and inertial measurements on a shared timeline.
For robotics researchers, those streams make it possible to compare what an RGB image suggests with the sensor’s range measurement. The card provides calibration and one MCAP recording per episode. It documents several encodings and invalid readings, especially on dark or reflective surfaces. The current payload is about 3.4 TB, and the dataset listing specifies a Creative Commons Attribution-NonCommercial 4.0 license.
Root Access Take · A depth sensor gives you another measurement to question. Check where points disappear on metal, glass, or a hand in motion before using a recording as a reference. Keep that uncertainty attached to the data when you pass it to someone training a model.
Try next · For a noncommercial experiment, pick one episode using the metadata. Open the RGB frame and range image side by side, then mark three places where the range reading is missing or ambiguous. Compare those same points across frames before planning a larger download.
Sources · BitRobot dataset card and license · Creative Commons BY-NC 4.0 terms
Explore the Root Forward archive, AI Radar, Governance guide and The Root Shelf.