ROOT FORWARD / NOVEMBER 8, 2026
Where AI is going—and what you can do with it.
An ongoing agent gets a job, a development kit firms up its foundations, and stale context becomes an operational concern. Faster image generation and more explicit compute-campus commitments bring the same question into view: what does the whole service need to work well?
An ongoing AI job needs a clear boundary, a current source of truth, and a human who owns the outcome.
Root Note
01 / CAPABILITIES / HUMAN–AI COLLABORATION
Perplexity gives ongoing work a named agent
Perplexity’s November 6 changelog highlights Teammates, agents assigned continuing responsibilities rather than a fresh prompt for every task. The help center describes planning, parallel work, persistent project context and monitoring of connected tools. A teammate can bring a result back for review and ask for input when it needs direction.
Access has an important boundary. The initial research preview is for individual Max subscribers on the web. Enterprise Pro and Enterprise Max are not included by default. Teammates inherit relevant plugin and Project permissions, and complex or frequent work consumes credits. A familiar workplace role does not establish that the service is approved for your organization’s customer data.
Root Access Take · The interesting design change is the job description. An ongoing responsibility needs a definition of done, a reporting cadence and a point where the agent must stop and ask. Otherwise the helpful new colleague can turn a small request into a permanent department.
Try next · Write a role for a weekly public-source vendor briefing before connecting any accounts. Specify the sources, three questions to answer, maximum scope and review step. Compare its second report with its first: did memory improve continuity, or simply preserve yesterday’s assumptions?
Sources · Perplexity’s dated changelog · Preview access, permissions and usage
02 / INTEGRATIONS / AGENT DEVELOPMENT
Google ADK graduates more of the machinery agents depend on
Google’s ADK Python 2.14.0 package appeared on PyPI on November 7; its release notes carry a November 4 version heading. The release promotes execution environments, A2A integration, evaluation and other runtime modules out of experimental status. That gives developers a clearer supported foundation for composing agents, tools and execution environments.
The same release adds repeated evaluation runs and configurable parallelism, allowing a case to produce a pass rate across several attempts. It also adds semantic matching for tool arguments and a Microsoft Teams runner. These are framework capabilities: a working deployment still needs the appropriate dependencies, account configuration, models and application permissions. An upgrade also deserves a migration review, including changed enum handling in function tools.
Root Access Take · Stable plumbing matters because service teams need something they can maintain after the demo’s author goes on holiday. Repeated evaluation is especially useful for agents whose first successful run hides inconsistent behavior. A green result becomes more convincing when it survives several ordinary attempts.
Try next · Pin 2.14.0 in an isolated development environment and replay one existing workflow several times. Record tool arguments, completion state and variation between runs. Review the documented breaking changes before changing a shared environment, and keep a known working dependency set for comparison.
Sources · Google’s release notes and migration details · Published package and requirements
03 / MSP PRACTICE / CONTEXT STEWARDSHIP
Google measures the operational cost of stale context
Google engineers Sam Saccone and Madhuri Garikapati published an analysis of agent distraction on November 6. They used traces from internal coding work to examine how old inline references could pull agents away from the user’s requested task. The broader finding concerns the material an agent reads while deciding what work belongs in scope.
The report connects unrequested detours with additional steps, tokens and waiting time, using statistical controls for several workflow characteristics. It also describes a cleanup process that checked historical intent before proposing removal of obsolete references for human review. These are observations from Google’s environment, rather than a universal forecast of savings from deleting documentation.
Root Access Take · Knowledge maintenance now affects execution. For an MSP, an obsolete onboarding note can compete with the current runbook if both arrive as apparently authoritative context. The useful cleanup is to establish which instruction still applies and who owns it. Merely making the knowledge base shorter can remove the explanation the next engineer needs.
Try next · Pick one frequently used service runbook. Mark its owner, review date and authoritative procedure, then inspect the older documents it links to. Ask an assistant to explain which instructions govern a sample task. Resolve contradictions with the human owner before letting that material drive an automated change.
04 / MULTIMODAL AI / INFERENCE HARDWARE
NVIDIA shows the tradeoffs behind faster FLUX image generation
NVIDIA’s November 6 tutorial walks through FLUX 3 Image Klein inference in TensorRT LLM VisualGen. It combines compilation, lower precision, attention optimizations and caching, with measurements on Blackwell hardware. The practical contribution is a documented path from an image model to a configured generation service, including examples of how optimizations change the output.
The reported tests use specific GPUs, resolutions and checkpoint variants. NVIDIA also states that its latency measurements exclude prompt upsampling, a separate step Black Forest Labs recommends for ordinary short prompts. The model card lists the FLUX Kommunity License, with conditions on commercial use. Weight access, runtime support and permission to use the output in your business each need checking.
Root Access Take · The fastest kernel is only one part of the service. A client waits for prompt preparation, generation, transfer and any correction pass. A small visual difference can be harmless in a mood board and expensive in a product image. Define acceptable output before choosing the speed setting.
Try next · If you have suitable hardware and model access, compare one supported baseline with one optimization using the same prompts and references. Measure total turnaround, inspect the images and record the exact configuration. Include the prompt-expansion stage in the budget, even when the published benchmark omits it.
Sources · NVIDIA tutorial, hardware and measurement scope · BFL model card, prerequisites and license
05 / GOVERNANCE / INFRASTRUCTURE ACCOUNTABILITY
OpenAI puts compute-campus promises into a community compact
On November 5, OpenAI published a National Community Compact and a Compute Campus Policy Blueprint. The compact describes commitments for future campuses OpenAI owns and operates, including local engagement, energy and infrastructure costs, resource reporting and workforce planning. The blueprint separately asks U.S. policymakers to establish broader rules for the industry.
The distinction matters. OpenAI issued these documents as company commitments and policy proposals; they do not establish a new certification standard or enacted nationwide requirement. The compact includes site-level reporting on water and electricity use and public progress reporting. Its stated scope also matters when evaluating a facility operated by another organization or under a different arrangement.
Root Access Take · A promise becomes useful procurement evidence when someone can identify the responsible entity, measurement, reporting date and response to a shortfall. That habit scales down nicely to an MSP choosing an AI supplier. An attractive sustainability page should lead to evidence that applies to the service being purchased.
Try next · Add four fields to an AI vendor review: who operates the infrastructure, which public commitment covers it, where results will appear and when to check again. Keep proposed policy separate from contractual obligations and applicable local requirements. Revisit the record when actual site reports arrive.
Sources · OpenAI’s dated announcement and scope · National Community Compact · Compute Campus Policy Blueprint
Explore the Root Forward archive, AI Radar, Governance guide and The Root Shelf.