OpenAI and Google moved agents from personal prompts into shared infrastructure. Workspace Agents added schedules, memory, and organizational audit data; Google folded identity, runtime, and observability into one cloud platform. A local privacy classifier and separate training/inference TPUs show the same specialization lower in the stack.
1. Workspace Agents turn team procedures into scheduled cloud work
OpenAI launched Codex-powered Workspace Agents for Business, Enterprise, Edu, and Teachers plans. A team can build, schedule, and share an agent through ChatGPT or Slack, with persistent files, tools, and memory in a cloud workspace.
The product turns a prompt into a maintained organizational asset. OpenAI's Compliance API records configuration changes and runs, while administrators can suspend agents or limit builders, sharing, tools, and actions. Credit pricing after May 6 also makes every schedule a recurring compute commitment.
Sources: OpenAI on Workspace Agents · Workspace Agents for business
2. Privacy Filter runs 128K-token PII labeling locally
OpenAI released Privacy Filter under Apache 2.0 with 1.5 billion total parameters, 50 million active parameters, and a 128,000-token context. The bidirectional classifier labels eight PII categories in one pass and can run locally before data reaches a remote model.
On OpenAI's corrected PII-Masking-300k benchmark, the model scored 96.79% precision and 98.08% recall; both the correction and scoring came from OpenAI. The architecture makes the operating point explicit, so high-sensitivity uses can favor recall and accept more manual review instead of hiding that tradeoff inside one F1 score.
Sources: OpenAI Privacy Filter
3. Google splits TPU 8 into training and inference variants
Google announced TPU 8t for model training and TPU 8i for inference. The split assigns large shared-memory pools to training while 8i targets responsive, multi-step agent workloads instead of forcing both profiles onto one accelerator design.
The launch article supplies positioning but no reproducible workload behind the headline speed and cost claims. The strategic move is narrower and clearer: Google Cloud can tune its own silicon by phase while continuing to sell Nvidia hardware, expanding procurement choice inside one cloud rather than replacing GPUs outright.
Sources: Google on TPU 8t and 8i · TechCrunch on the two-chip strategy
4. Gemini Agent Platform makes identity persistent across multi-day runs
Google launched Gemini Enterprise Agent Platform as the evolution of Vertex AI. It combines a multi-day runtime, memory bank, cryptographic agent identities, registry, gateway, simulation, evaluation, and observability across more than 200 first-, third-party, and open models.
The cryptographic identity is the key distinction from a shared service account. It can bind an action to one agent and policy even after a workflow spans days or delegates to subagents. Consolidation also gives Google control of the execution record, making trace export and identity mapping the portability test for existing Vertex deployments.
Sources: Google Cloud on Gemini Enterprise Agent Platform · Google Agents CLI repository