The industry spent this week doing two things at once: shipping production-grade agent infrastructure and quietly retiring the services that came before it. Microsoft and Google both pushed serious agent tooling to general availability, AWS moved a swath of its early AI portfolio to maintenance mode, and the UK Treasury formally brought Google Cloud under financial regulation. The scaffolding around agents is being built and dismantled in the same breath.
Microsoft ships the agent platform it promised at Build
Microsoft moved a coordinated set of Foundry capabilities to general availability: OpenAI’s GPT-5.6 series (Sol, Terra and Luna, priced from $1 to $5 per million input tokens), hosted agents in Foundry Agent Service, toolboxes for dynamic tool selection, and publishing direct to Teams and Microsoft 365 Copilot. A new Asia-Pacific Data Zone keeps frontier model processing within the region, and the GPT-5.6 models land in all 28 global regions from day one.
Why it matters: Model access is no longer the differentiator; the runtime, governance and distribution layer wrapped around agents is where the real platform lock-in will happen, and Microsoft just shipped its version of that layer.
Google gives untrusted agent code somewhere safe to run
Cloud Run sandboxes entered public preview: isolated execution environments that start in milliseconds inside your existing Cloud Run instances, with zero network egress by default, no access to credentials or the metadata server, and a read-only filesystem overlay. Google demonstrated 1,000 sandboxes starting, executing and stopping with an average latency of 500ms. Because sandboxes share the resources of the running instance, there is no additional charge.
Why it matters: Isolation for AI-generated code just became a platform primitive rather than a bespoke microVM project, and the zero-cost pricing puts immediate pressure on the specialist sandbox vendors.
Your agent benchmarks may be lying to you
Google’s Data Cloud frontier AI team published a methodology piece on Discovery Bench, a framework that varies the ambiguity of evaluation queries using information theory rather than pass or fail scores. The findings are uncomfortable: one agent scored a perfect F1 on a query at neutral phrasing and zero once a single distinguishing token was removed, and a widely used public benchmark turned out to contain broken ground truth that other teams had already built on.
Why it matters: If you are gating agent deployments on benchmark scores, a passing grade tells you nothing about how close the agent sits to a cliff, or whether the ruler itself is broken.
AWS retires its first generation of AI services
Buried in the AWS weekly roundup was the quarterly service availability update, and it makes sobering reading. Amazon Kendra, Amazon Q Business, Bedrock Agents (now rebranded Classic) and a long list of SageMaker features move to maintenance mode, closed to new customers from 30 July. The same roundup announced Claude Sonnet 5 on AWS and general availability of Amazon WorkSpaces for AI agents, so the replacements arrived in the same post as the retirements.
Why it matters: Q Business launched barely two years ago, so if it sits anywhere in your estate you now need a migration plan, and if it does not, the shelf life of first-generation AI services is worth pricing into your next adoption decision.
The Bank of England now oversees Google Cloud
On 10 July the UK Treasury designated Google Cloud EMEA as a critical third party to the UK financial sector, placing it under direct oversight by the Bank of England, the PRA and the FCA. The designation reflects the number of UK firms running material workloads on Google Cloud, and follows the critical third party regime the regulators established in 2024.
Why it matters: For UK financial services architects your cloud provider is now supervised alongside you, which strengthens the sector’s resilience posture but shifts none of your own obligations under SS1/21 and SS2/21.
The launches this week all assume agents will run in production at scale, and the retirements show what happens to services designed before anyone knew what that would look like. It is worth asking which of this week’s general availability announcements will survive their own first-generation moment. The gap between shipping an AI service and knowing whether it was the right shape has never been shorter, or more expensive to get wrong.
