Your agent problem is now a governance problem

Three things happened this week. They all say the same thing: the agent experiment phase is over. The governance phase started.

1. Salesforce flipped Agentforce to default-on. Every eligible org gets it automatically — the Setup opt-in is gone. It’s access management now, not a feature gate.

Here’s what that means for a service team: you woke up with a dormant agent capability sitting in your org. Nobody filed a change request. Nobody ran a risk review. It’s just there, waiting for someone to grant it access. That’s a new attack surface and a new audit line that arrived without anyone’s permission.

The field lesson: the model was never the risky part. The wrapper is. Who can invoke the agent, what data it can touch, what it’s allowed to write — those are deterministic decisions, and they need to be made on purpose. If your agent governance was “we’ll figure it out when we turn it on,” default-on just made that decision for you.

2. Dreamforce stopped selling blank agents and started selling job roles. Seven role-specific agents, long-horizon execution, multi-agent orchestration, and an Agent Optimizer that auto-tunes performance.

The market moved from “build an agent” to “staff a job.” That’s the right framing, and it’s how the serious production deployments already work: scope the role, define what good looks like, regression-test it the way you’d performance-review an employee. The Agent Optimizer is the vendor catching up to what good teams already do with golden test sets and release gates.

3. A competitor shipped the article-coach concept. A ServiceNow-native vendor launched an agentic Knowledge Article Optimizer: finds failing searches, scores articles field-by-field against standards, generates optimized titles, metadata, and bodies.

The idea is validated — on a rival platform. The knowledge layer is the moat. Agents are only as good as the articles they ground on, and most teams’ knowledge bases are full of duplicates and quiet contradictions. No model upgrade fixes that. Only the wrapper does — a draft check that catches contradictions before they ship.

The takeaway: design the system around the AI — deterministic wrapper, probabilistic core — because the vendors just made the wrapper your problem.

One thing to try this week: pick three pairs of knowledge articles you suspect contradict each other. Ask your agent one question each pair answers. Count how often it picks the wrong one, blends both, or stays silent. If it can’t detect a contradiction, your draft-review checklist is missing its most important check.