First victim of AI: voice verification. I’m glad Verizon is doing this but was really hoping banks would be the first to remove this feature. Worth contacting yours to make sure voice ID is off for your account.

In the same week, Microsoft, AWS, Google, and Mistral all shipped variants of the same primitive. The category-defining unit of agent infrastructure has quietly shifted from "model plus tool loop" to "stateful runtime that survives time boundaries." Satya on Foundry: "durable, stateful agents that run across time boundaries, orchestrate tools and models, and close the loop with evaluation and improvement over long-running workflows." Andy Jassy, announcing OpenAI on Bedrock, slipped in the actually interesting part: "alongside the upcoming Stateful Runtime Environment." Bedrock is no longer trying to be a model gateway. It's trying to be the place an agent lives between turns. Google Cloud's Gemini Enterprise pitch this week was almost identical: "long-running agents" that handle end-to-end financial reconciliation or multi-week sales sequencing, "from ad hoc to recurring tasks." Mistral shipped Workflows, called explicitly "the orchestration layer for enterprise AI." Four hyperscalers and one frontier lab, same week, same primitive. Why this is the right abstraction: The single-shot chat completion was a bad fit for real work. Real work is checkpointable, multi-turn, multi-day, multi-agent, sometimes paused for human approval, sometimes resumed against new evidence. Building that on top of stateless model calls means every harness reinvents the same machinery: durable storage, retry logic, approval interrupts, partial-failure replay, audit trails, cost accounting per agent run. Everyone was building it. Now the cloud is offering it. What "stateful runtime" actually has to do well: Persistence that survives the model. The agent's working memory and intermediate artifacts have to outlive a single inference call and a single VM. That means a real datastore with versioning, not a Python dict. Time-aware scheduling. "Run again when the invoice posts." "Wait three hours, then check." "Pause until a human approves." The runtime owns the clock. Pluggable models. The runtime can't be locked to one model vendor or it dies the day a better model ships. AWS clearly understood this, which is why they shipped Stateful Runtime and OpenAI on Bedrock in the same breath. Audit and observability per run. Compliance and debuggability are the same problem at scale. You need to replay any agent step with the exact tools, prompts, and state it had at that moment. Cost accounting per agent identity. Otherwise FinOps cannot survive an agent fleet. The competitive question is who owns the runtime, not who owns the model. The model vendors all know this, which is why Mistral is shipping Workflows directly and OpenAI's Codex/AgentKit/Apps stack keeps creeping toward "platform." If the runtime is on AWS, AWS gets the lock-in. If it's on the model provider, the model provider gets it. Whoever owns the place agents live between calls owns the relationship. For builders: stop hand-rolling agent state. The next twelve months are going to be a Cambrian explosion of "durable workflow" SDKs from every major cloud, and the cost of building your own orchestration layer just went above the cost of adopting one. Pick the runtime your eval harness can target portably and move on. The interesting work is upstream of plumbing now.