$ today_in_10_seconds
Amazon is leaning on its own engineers to stop wasting CPU cycles — because agentic workloads have made ordinary processors the scarce resource, not GPUs. Meanwhile MCP shipped its biggest rewrite ever, AWS quietly put Bedrock Agents out to pasture, and the EU AI Act's enforcement powers went live. Four stories, one theme: the agent stack is being re-plumbed for production.
$ lead_story
Amazon tells engineers to stop wasting CPU as the agentic crunch bites
AWS has asked internal engineering teams to cut CPU waste in order to protect capacity, per a report from The Information that surfaced around August 3. The reason is structural, not budgetary: agent workloads are CPU-native in a way training never was. Intel told investors that data center CPU-to-GPU deployment ratios have already tightened from 1:8 to 1:4, and could converge on 1:1 as workloads shift from training to agentic inference — because the steps in an agent loop that aren't inference (breaking down goals, navigating links, parsing files, making network calls, executing code) are all CPU tasks. The supply chain planned for a GPU-dominated world and is now short: server CPU lead times have stretched from roughly two weeks before the agentic spike to about six months.
// why_it_matters: Your agent's cost curve is set by the boring compute you forgot to model, not the tokens you obsess over.
$ also_worth_attention
MCP ships its largest revision since launch
The 2026-07-28 spec landed on July 28 with a stateless protocol core, multi-round-trip requests, header-based routing, cacheable list results, authorization hardening, and a formal extensions framework. The initialization handshake and Mcp-Session-Id are gone, which means servers no longer need sticky routing or shared session state behind a load balancer — and Roots, Sampling, Logging, and the legacy HTTP+SSE transport are all deprecated on a twelve-month window. The insight: this is MCP admitting agents are high-fanout infrastructure, not long conversations.
link →AWS retires Bedrock Agents in favor of AgentCore
Bedrock Agents is now Bedrock Agents Classic, closed to new customers since July 30 and in maintenance mode — and its model catalog is frozen as of that date, so new models will only be reachable through AgentCore. Existing agents keep running, but a frozen catalog is a hard ceiling on capability. If you prototyped on Classic, your migration clock started a week ago.
link →The EU AI Act's teeth came out on August 2 — but not the ones you were told
Article 50 transparency obligations became generally applicable and enforceable across the EU on 2 August, covering any product that talks to users or generates images, audio, video, or text — regardless of high-risk classification. GPAI enforcement powers and the full penalty regime also switched on, while the heavy high-risk obligations were pushed to 2 December 2027 and 2 August 2028. Translation: the deadline you prepared for moved, and the one you ignored arrived.
link →
$ bigger_picture
The trend behind the headlines
Every story today is about operability, not capability. The compute layer is repricing around CPU-bound agent loops. The protocol layer just threw out sessions so agents can scale on ordinary HTTP. The platform layer is re-architecting from single-agent tool-calling to systems of agents with shared memory and gateways. And the governance layer started issuing fines. Gravitee's research this week puts the gap plainly: more than 7 million AI agents are already running inside large enterprises, more than double six months ago — yet only around 11% of enterprises run agents in production. Capability was never the constraint. Knowing what your agents did, what they cost, and who authorized it — that's the constraint.
Langfuse
Open-source LLM and agent observability: per-run traces, token and cost attribution, evals, and prompt versioning. If today's lead story landed uncomfortably, this is the cheapest way to find out which agent loop is quietly burning your budget before finance does. Self-hostable, which matters if your traces contain anything you can't ship to a vendor.
$ subscribe --free
Get the next edition in your inbox
// DAIly Agent covers agentic AI three times a week
// free, no spam, exit anytime