Agent governance is the bottleneck no one is measuring
Three teams describe the real agent problem: 5 agents per employee with no oversight.
Start with monitoring and identity before you add controls or sponsors.
Agent sprawl is real. Runlayer found 300 agents for 60 people, most unattended.
Cost spikes from a vibe-coded MCP server went unnoticed until the bill arrived.
Without registry and risk scoring, you can't govern what you can't see.
Harness trumps model. Arize's agent quality gains came from harness design, not waiting for a better model.
Anthropic's Rust rewrite saved 40% CPU without any model change.
Code mode cut tokens 43% by replacing tool lists with a sandboxed environment.
Inference cost is falling. The viral $132M token spend claim was off by 20x after accounting for cache hits.
Per-task cost for fixed intelligence dropped 58x in a year.
But teams will reinvest savings into harder tasks, so budget for expanding ambition.
My take
I buy the monitoring-first approach over the agent-as-employee analogy.
Naming a human sponsor for every ephemeral agent won't scale past a few dozen.
What will scale is a registry, intent-based permissions, and a cost dashboard.
The inference cost trend is real, but don't assume teams will use the cheap old model: they'll chase smarter agents for harder work.
Start with visibility, then layer controls as you learn where things break.
The videos behind this brief
We watched these so you don't have to.
300 agents in 60 people: The ungoverned AI agent sprawlAI Engineer
Why Agent Governance Needs Brakes, Not WallsAI Engineer
Why Your AI Agents Need a Manager, Not Just PermissionsAI Engineer
Agent to-do lists waste tokens. Move the plan to the harness.AI Engineer
Anthropic's Efficiency Is a Product of Loops and Rust, Not Just aLatent Space
Code Mode Agents: Less Token Bloat, More Sandbox ComplexityAI Engineer
Inference costs are crashing — don't let bad math scare you offTheo - t3․ggVideo and thumbnails remain the property of their respective creators. tldw.news provides editorial analysis, commentary, and discovery links to original content.