Good morning. Here is what matters in AI today, and how to put it to work.
AWS ships Dogwood and DynamoDB vector search the same week, giving teams two missing pieces for production-ready agentic AI.
~3 min read · last 12 hours
In today's issue
01
AWS open-sources Dogwood to govern sequences of agent tool calls
02
AWS adds native vector search to DynamoDB
03
Google Pixel 11 Pro vs. Apple iPhone 17 Pro: a tight race at the top
04
Google Pixel 11 vs. Pixel 9: the upgrade case is stronger this year
Main story
AWS open-sources Dogwood to govern sequences of agent tool calls
Dogwood extends the Cedar policy language with temporal conditions, so rules can reason about the order and timing of actions an AI agent takes, not just whether a single action is permitted.
Why it matters: Teams deploying multi-step agents have had no standard way to enforce policies across a chain of tool calls; Dogwood fills that gap and is worth evaluating before your next agentic release.
What to watch next: Watch for other cloud providers to respond with their own temporal policy primitives for agents, and for Cedar-compatible tooling to emerge from the open-source community once Dogwood gains adoption.
We see two foundational moves this week that together push agentic AI closer to production readiness: AWS is giving teams a formal way to govern what agents are allowed to do over time, while also making it easier to store and query the vector embeddings those agents depend on, all inside existing infrastructure.
The hardest unsolved problems in agentic AI are governance and data access, and AWS moved on both fronts simultaneously this week. Dogwood gives engineering teams a principled, auditable way to constrain what an agent can do across a sequence of steps, not just at a single decision point. Native vector search in DynamoDB means teams can add semantic retrieval without bolting on another managed service. Together, these moves lower the operational and compliance barriers that have kept many agentic workloads in the prototype stage. On the hardware side, the Pixel 11 confirms that on-device AI is now a baseline expectation, not a differentiator, which reshapes where product teams should plan to run inference.
All the best, the KYFEX team
Quick hits
Agentic AI gets guardrails and data infrastructure
AWS adds native vector search to DynamoDB
DynamoDB can now store embeddings alongside application data and run vector similarity queries natively, removing the need for a separate vector database in many RAG and semantic-search architectures.
Why it matters: If your team already runs DynamoDB, this cuts one service from your AI stack and simplifies the operational surface for embedding-based features.
Google's new Pixel lineup shows how aggressively on-device AI capabilities have become the primary battleground in consumer hardware, putting practical pressure on enterprise mobile strategies that assumed cloud-only inference.
Google Pixel 11 Pro vs. Apple iPhone 17 Pro: a tight race at the top
Hands-on testing shows the two flagship phones are closely matched, each with distinct strengths, signalling that on-device AI differentiation is now table stakes rather than a premium differentiator.
Why it matters: Product teams building mobile AI features need to account for near-parity hardware capabilities across Android and iOS when deciding where to run inference.
Google Pixel 11 vs. Pixel 9: the upgrade case is stronger this year
For users who skipped the Pixel 9 and 10 generations, the Pixel 11 offers enough accumulated AI and hardware improvements to make the jump worthwhile.
Why it matters: The pace of on-device AI progress means even a two-year-old flagship is now meaningfully behind, a signal for enterprise device-refresh planning.
Draft an agent policy in plain language before writing code
I am building an AI agent that can call the following tools: [list your tools here]. For each tool, describe: (1) under what conditions the agent is allowed to call it, (2) any ordering constraints (e.g. tool B must not run before tool A has succeeded), and (3) any rate or time-window limits. Output a numbered policy list in plain English that I can later translate into a formal policy language.
Why it helps: With AWS Dogwood now available to encode temporal agent policies, drafting them in plain language first makes the translation to formal rules faster and surfaces edge cases before they reach production.
Before you ship it
The risk
Agentic systems that chain tool calls can accumulate permissions over a session, meaning a policy that looks safe for a single action may allow harmful sequences when combined across steps.
Do this
Define and document the full expected sequence of tool calls for each agent workflow before deployment, then verify that your policy layer (such as Dogwood) explicitly rejects any sequence that falls outside that specification.
Ready to ship AI, not just read about it?
KYFEX designs and builds production AI for teams that need it working, not just demoed. Tell us what you're working on and we'll bring the engineering.
