When You Tag Claude in Slack, That's One Kind of Agent. When You Don't, That's Another.

Most enterprise AI governance conversations assume the agent was triggered by a human. Someone typed a prompt. Someone clicked a button. Someone @-tagged the model in a thread.

That assumption is increasingly wrong.

Two modes. One platform. Zero distinction.

When an employee tags @Claude in a Slack thread, something specific happens: there is a human intent signal, a bounded context window, a clear start and end, and a visible record in the thread. The interaction is reactive. The agent responds and stops. The governance question is relatively contained: did the response comply with policy?

When an enterprise deploys a Claude agent with memory, tool access, and a standing scope of work — with minimal or no human prompting per session — something fundamentally different happens. The agent runs for hours. It accumulates context across sessions. It takes actions outside of any single conversation thread. Nobody @-tagged it. Nobody is watching the specific moment it makes a decision.

These are not variations of the same governance problem. They are different governance problems entirely.

The reactive mode

A reactive tagged session has high signal density. The human intent is explicit. The context window is bounded. Deviations tend to be abrupt and visible.

For governance, this means: shorter persistence windows matter. A deviation that lasts 20 minutes is meaningful. The evidence for intended variation is strong when it exists — if a policy changed and the agent responded differently, you can trace the causation.

Most enterprise AI governance tools were designed for this mode. They assume a prompt-response cycle. They look at individual sessions. They flag anomalies in single interactions.

The ambient mode

An ambient proactive agent doesn't get @-tagged. It was given a standing scope: monitor this channel, summarize this data, maintain this workflow. It operates with scoped memory and tool access. It may run for days between meaningful human check-ins.

For governance, this mode is categorically harder:

• Drift is slower. A 30-minute deviation is noise. A 12-hour shift in cost patterns is a signal.
• Intent is less explicit. When no human triggered the session, "intended variation" is harder to prove.
• External side effects are higher risk. An ambient agent with tool access may have written files, sent messages, or modified state that no single human approved or reviewed.
• The audit trail is thinner. There is no Slack thread to point to. There is no @-tag timestamp. The evidence of what happened has to come from the governance layer itself.

This is the mode that Article 14 of the EU AI Act was written for. Human oversight of long-running autonomous systems is not optional when the system is making consequential decisions without explicit per-session human authorization.

What PromptKing does differently

The behavioral baseline fingerprinting and drift classification systems that shipped in v4.4.0 and v4.4.1 treat these two modes separately — by design.

Reactive tagged sessions form baselines over shorter observation windows. Drift thresholds are tighter. Evidence requirements for intended variation are lower because human intent is visible.

Ambient proactive agents require longer observation windows before a baseline matures. Drift thresholds are wider — not because ambient agents are less risky, but because their natural behavioral variance is higher. When confirmed drift is detected on an ambient agent, the evidence requirements are stricter precisely because there is no explicit human prompt to point to as authorization.

The rollback primitive that shipped in v4.4.2 handles this distinction too. When an ambient agent has been operating with tool access, a rollback does not simply restore a conversation state. It flags all external actions taken during the drift window as manual_required — because those actions may have written state that cannot be cleanly undone, and the governance record must reflect that honestly.

The enterprise question nobody is asking yet

Most enterprise AI programs have a clear answer to: "What did our agents respond to this week?"

Almost none have an answer to: "What were our ambient agents doing between human check-ins — and did their behavior drift from what we authorized?"

That is the governance gap. And it is getting wider as more enterprises deploy standing AI workflows that don't require a human to trigger each session.

The difference between @Claude in Slack and an autonomous Claude agent running in the background is not just a product distinction. It is a governance architecture decision. Most enterprises have not made that decision explicitly yet.

PromptKing makes it explicit — and governable.