As AI agents enter workflows, the control layer grows
AWS opened a language that watches agent policies across a run and DynamoDB absorbed vector search; adoption in WeChat and workplaces shows agents settling rapidly into everyday infrastructure.
Artificial Intelligence··Evening
Dogwood watches an agent’s whole run
AWS open-sourced a policy language called Dogwood under the Apache 2.0 licence, InfoQ reports. Dogwood adds temporal conditions to AWS’s Cedar authorization language, so a rule can bound a sequence of tool calls read from an agent’s event history instead of judging one request at a time. The language introduces four operators: formerly, count_within, count_distinct_within and sum_within. AWS defines them as standard-library macros over a subset of a temporal logic. Any valid Cedar policy is also a valid Dogwood policy, so existing rules need no rewriting. AWS notes that the reference interpreter is for exploring and testing the language rather than running authorization in production. Policies that use temporal conditions also give up Cedar’s automated formal analysis. As agents make sequences of tool calls inside a workflow, the control layer has to move past approving a single request. Dogwood is presented as a language built for that gap: the policy view expands to the event history that accumulates across a run.[1]
DynamoDB absorbs vector search
Amazon DynamoDB now keeps embeddings alongside application data and answers approximate nearest-neighbour queries directly, InfoQ reports. The feature is generally available in every region where DynamoDB runs, and indexes support up to 4,096 dimensions. Distance can be measured with Euclidean, cosine or dot product, and queries can carry inline filtering. AWS reports single-digit millisecond latency at a scale of trillions of vectors. Billing sits on top of standard DynamoDB charges in three parts: data written to the index, data processed during searches and data stored. Esra Kayabali of AWS says vector indexes have no storage limits and scale horizontally as the data grows. Standing up a separate vector database meant an extra moving part for agent and retrieval workloads. DynamoDB’s absorption of vector search pulls that part beside application data. As agents need more memory and search, the storage layer sticks to the same infrastructure that now also carries temporal policy languages.[2]
Agent use grows in WeChat and at work
Tencent named Xiaowei for the first time in the quarterly results it announced on 14 August. The assistant is reachable in limited trial access, through two green dots at the top of WeChat’s main screen. Ben Jiang of the South China Morning Post handed it control for 24 hours. Xiaowei is built as an agent: it goes past recommending and books reservations and places orders through WeChat’s messaging, payments and mini-program layers. Tencent emphasises user privacy and inference efficiency. Access stays in limited trial, and the company has given no date for a broader rollout. On the same day, The Decoder reports an Epoch AI and Ipsos survey in which 20 per cent of employed US adults say they hand at least one task that used to go to a colleague or a contractor to AI instead. The survey ran on July 10-19 with 1,106 employed adults. Delegation is highest in software development at 57 per cent, followed by data analysis at 46 per cent. Respondents say they use 66 per cent of the AI output either unchanged or with only minor tweaks; against that, about one in six AI-assisted tasks now takes longer than it used to. The researchers read the result as a redistribution of tasks between people and AI, and they do not speak of full automation. While Dogwood and DynamoDB grow the control and search layers, the WeChat trial and the workplace survey show agents settling into everyday flows.[3], [4]
