Agent Ops: monitoring, hosting, stateless protocols, tooling, and modeling

Bloom Security emerges from stealth with $20M seed to monitor AI agents and extensions. The startup raises $20M to monitor AI agents, browser extensions, and endpoints from stealth. Outcome engineers must bake endpoint and agent observability into delivery pipelines — think of this as an emergent immune system and gate for agentic workloads (Principles 14, 15).

datasette-apps 0.2a0 ships invisible-iframe debugging and app-listing tools so agents can test and manage apps programmatically. That gives you a lightweight, auditable way for agents to exercise and inspect UI-level tools — useful for building reproducible agent artifacts and safer tool invocation (Principles 03, 07).

Runtime: MCP goes stateless; Baseten courts lab partners reports MCP’s shift to a stateless request/response core and Baseten’s Model Labs for closed-weight distribution. Stateless transport simplifies agent orchestration and retries, while Model Labs lowers the infra lift for serving proprietary models — both change deployment and graph design choices you’ll make for production agents (Principles 07, 11).

A deep dive into Nvidia’s Vera CPU and the Olympus cores that power it reveals a monolithic 88-core Armv9.2 CPU aimed at AI head-node and agent hosting at datacenter scale. Hardware tuned for dense agent hosting reshapes cost, failure domains, and how you partition agent work versus orchestration — plan for different placement and order assumptions (Principles 09, 12).

Explorative modeling: Train on the best of K guesses introduces an exploration axis to generative training that boosts sample, FLOP, and parameter efficiency across modalities. More sample-efficient generative training lowers the compute barrier for maintaining specialized agent models and speeds validation cycles for outcome-focused iterations (Principles 12, 16).