The New Bottleneck Isn’t Building Agents, It’s Governing Them

AI Agents

The New Bottleneck Isn’t Building Agents — It’s Governing Them

What happened: Akuity, the Kubernetes delivery platform, shipped an Agentic Control Plane and an MCP Server that give AI agents governed access to software pipelines. Instead of letting agents run loose with open-source MCP servers that carry no identity, no policy layer, and no audit trail, Akuity routes every agent request through existing pipeline controls — so an agent inherits the same permissions, identity, and audit logging as the human who connected it.

Why it matters: This is the clearest signal yet that the hard part of the agent era isn’t building the agent — it’s giving it a safe, reliable path to do real work. The early proof is striking: Major League Baseball’s DevOps team says that within hours of enabling the control plane, it surfaced more than one hundred degraded applications, then traced a systemic issue to a single root cause and had a fix ready in under ten minutes. That’s not an assistant suggesting things — that’s an agent operating across a fleet with guardrails.

What’s next: Watch for this pattern to repeat across every tool your agents touch. The winners won’t be whoever adopts agents fastest — Akuity CEO Hong Wang put it plainly: they’ll be whoever gives agents a governed route to production. If you’re building AI employees that touch real systems, identity inheritance and an audit trail aren’t nice-to-haves; they’re the thing that keeps you from a “rogue agent” headline. Source

TSA’s “Ace” Is Now Clearing 100K Traveler Questions a Month

The TSA’s Salesforce-built AI agent, Ace, handles roughly 100,000 traveler conversations monthly and resolves 96% of routine inquiries — packing liquids, clearing checkpoints with medical devices — without human escalation, cutting interaction costs by more than 90%. It’s a rare government deployment with published ROI, and a preview of what “deflect the easy stuff, escalate the hard stuff” looks like at national scale. Source

Voice AI Starts Selling on Outcomes, Not Minutes

SquadStack.ai is shifting its voice AI sales agents from per-minute pricing to outcome-based pricing — customers pay on revenue generated, not call volume. When a vendor prices on the outcome the agent produces, it’s a maturity signal: the tool is now confident enough in results to put its own margin on the line. Source

AI News

Gates Foundation Puts $1B on the Table to Close the Language Gap

The Gates Foundation will spend at least $1 billion over two years on AI access — including data work to make models function in languages they currently can’t. The incentive is stark: English is nearly half of all web content, while Chinese (1.3%) and Spanish (6%) trail far behind their speaker counts, leaving most of the world’s users with second-rate models. Source

Your AI Assets Are Now the Target, Not the Tool

Google’s Threat Intelligence Group reports adversaries are now going after proprietary models, API credentials, and cloud compute quotas — running “industrial-scale” distillation campaigns (100M+ prompts against Google’s models) and even deploying an autonomous multi-agent attack framework that harvested mass credentials in under six hours. The takeaway: API keys and service accounts are the new crown jewels, and abuse of them looks like legitimate traffic. Source

Medicare’s AI Prior-Auth Pilot Is Delaying Care, New Docs Reveal

Newly surfaced documents show Medicare’s WISeR pilot — which uses AI to review prior-authorization requests in six states — has caused care delays, work-flow disruptions, and added administrative strain, with one vendor approving 88% of requests instantly but flagging the rest for slow human review. It’s a cautionary tale for anyone deploying AI into a sensitive workflow: the automation is cheap, but the edge cases are where the harm lives. Source