ROOT FORWARD / NOVEMBER 15, 2026
Where AI is going—and what you can do with it.

MiniMax gives image edits more references to work from. Poppy gives personal agents a way to ask for access, Google Cloud expands agent monitoring, Figma and Stanford test whether generated graphics actually move as requested, and llama.cpp gives application developers a clearer upgrade boundary.
A handoff is only as clear as what each side can verify.
Root Note
01 / MODELS / IMAGE EDITING
MiniMax H3-Image puts reference editing in the same model
MiniMax released H3-Image on November 11. The same image model generates from text, builds a new composition from references, and edits an existing image. MiniMax documents up to 16 references and output up to 4K on its hosted design surface. More references let you assign the subject, layout, and palette to different inputs. They also give you more ways to supply conflicting instructions.
The model card lists roughly 69 GB for the image package and another 67 GB for its separate text encoder. Those are download sizes. Runtime requirements depend on the serving setup, so consult the documented accelerator configurations before reserving hardware. The downloadable weights carry a custom community license. Its commercial terms require prior written authorization for providers of third-party hosted services or when the defined revenue threshold is exceeded. Check the actual terms before packaging this for clients.
Root Access Take · A reference is useful when you can tell what it constrained. Give each input a job and decide which parts of the original must survive an edit. Otherwise, choosing a winner turns into judging which picture looks expensive.
Try next · Use a fictional product label. Ask the model to change the lighting while preserving the words, proportions, and layout. Score each of those separately across several attempts.
Sources · MiniMax announcement · Model card and deployment notes · Community license
02 / INTEGRATIONS / AGENT ACCESS
Poppy 1.0 gives a business a way to greet a personal agent
Sierra announced version 1.0 of the Personal Agent Protocol on November 13. The protocol gives a personal agent a way to discover the interfaces a business exposes, identify itself, and request delegated account access. A business publishes a discovery document that can point to a company agent or an API. Account access uses OAuth, so the customer can approve access and later revoke it.
Sierra reports that Shopify now publishes Poppy descriptors for merchants and that Meta’s Muse looks for them. The Shopify example also shows the division of work. Poppy handles discovery and access, while the Universal Commerce Protocol handles shopping actions such as populating a cart and checking out. What a customer can actually do still depends on what the business exposes, what the agent implements, and what access the customer approves.
Root Access Take · The useful product decision is where to draw the permission boundary. Give order tracking, address changes, and cancellations different approval requirements. The service owner can then explain and test what each grant actually permits.
Try next · Start with a fictional store and one read-only order-status tool. Check the same request signed out, after explicit approval, and after revocation. Write down what the agent tells the customer at each step.
Sources · Sierra release and partner examples · Poppy 1.0 specification
03 / MSP PRACTICE / OBSERVABILITY
Google Cloud gives support teams more ways to trace data agents
Google Cloud expanded observability for data agents on November 13. BigQuery Studio’s generally available monitoring dashboard tracks Conversational Analytics usage, sessions, knowledge sources, and latency. Looker’s generally available System Activity dashboards track engagement and token consumption. Ratings and query review in Looker’s Responses & feedback tab are in preview.
For teams that need detail about an individual turn, BigQuery can write prompts, responses, and telemetry to a customer-owned table when conversation logging is enabled. Google documents conditions where logging is skipped, including when a knowledge source uses row-level security or dynamic data masking. Custom agents have their own instrumentation path. Those differences matter when two agents look healthy on the same dashboard but give support staff very different evidence to work with.
Root Access Take · Decide what a ticket needs before turning on more logging. A request ID, source name, latency, and outcome can speed a handoff. Retaining every prompt or response broadens who can see that content and adds another retention obligation.
Try next · Run one synthetic support question through a test agent. Ask an on-call teammate to find the slow turn, identify the source it used, and explain any expected gaps using only the access they would normally have.
Sources · Google Cloud launch and feature status · BigQuery conversation logging documentation
04 / EVALUATION / ACCEPTANCE CRITERIA
MotionGraphicsBench turns an animation brief into checks
Figma and Stanford researchers introduced MotionGraphicsBench on November 10, following a November 7 preprint. Its 100 tasks each give a model a static SVG and a motion brief. The researchers evaluate the rendered frames against requirements such as the timing of a movement or the changing position of two objects. A clip can render successfully and still fail because it does the wrong thing at the wrong time.
The team reports that its verification programs agreed with human judgments on 93.3% of 180 held-out animations under its study design. That figure describes agreement in this benchmark. Use it as an example of making a check repeatable. The NIST AI Risk Management Framework is voluntary guidance, ISO/IEC 23894:2023 provides risk-management guidance, and ISO/IEC 42001:2023 specifies requirements for an organizational AI management system. Each has a broader purpose than scoring a generated artifact.
Root Access Take · Write the acceptance criteria when you write the brief. A timing check can give reviewers a repeatable answer. Keep human review for judgments such as whether the animation makes the information easier to follow.
Try next · Take one generated chart animation and check that its labels remain legible, each legend entry appears before its series moves, and the final frame matches the source values. Have a reviewer assess whether the sequence reads clearly.
Sources · Figma and Stanford research walkthrough · Research preprint · NIST AI RMF · ISO/IEC 23894 · ISO/IEC 42001
05 / LOCAL AI / LIBRARIES
llama.cpp and ggml give downstream developers a version boundary
The maintainers of llama.cpp and its underlying ggml library announced version 1.0.0 on November 13. Future breaking changes to the public C interface or binary compatibility will get a major-version bump. Compatible feature additions, including new model or backend support, can ship in a minor version. That gives applications which embed these libraries a clearer signal about the kind of change they are taking on.
For a support team maintaining a local inference stack, the integration boundary depends on how the application uses it. An executable you call from a script has a different dependency surface from a shared library you link into an application. Record the package distribution, hardware backend, model file and library version together. A working install on a developer’s laptop tells you little about whether the same client build will load the right backend on the deployment machine.
Root Access Take · Version numbers help you route an upgrade to the right reviewer. Make the support record specific enough that a second person can rebuild the same combination. Start with the version behind your last good deployment.
Try next · Pin the known-good build and the candidate on one representative machine. Use the same model file and backend, exercise startup and one fixed prompt, then confirm that reverting the package restores the baseline.
Sources · Maintainers’ release and versioning notes · llama.cpp release · ggml release
Explore the Root Forward archive, AI Radar, Governance guide and The Root Shelf.