No Wild Agents: The Case for Controlled Delegation

The phrase "AI agent" is becoming too broad to be useful.

Some people mean a chatbot with a tool. Some mean a browser automation loop. Some mean a long-running worker that can touch files, email, databases and calendars. The risk is that everything gets marketed as autonomy before the control layer is mature.

For real work, especially in Switzerland, uncontrolled autonomy is the wrong default.

The Better Primitive

The better primitive is not an all-powerful agent. It is a bounded worker:

• one role
• one workspace
• a limited tool list
• explicit permissions
• visible cost limits
• durable logs
• approval gates before external action

This is less exciting than a demo where an agent "does everything". It is also much closer to what businesses can safely buy.

One Customer, One Workspace

For customer-facing systems, every customer should get an isolated workspace.

one customer or coach organization
-> isolated workspace
-> fixed agent team
-> separate data, credentials, logs and budgets

This matters because data boundaries are product boundaries. If a global agent can see every customer, every failure becomes a catastrophic failure.

The Agent Team

A useful business workflow usually needs multiple small agents:

• scout: collect source information
• scorer: evaluate fit
• drafter: prepare text
• reviewer: critique with fresh context
• proof capturer: record evidence
• operator: summarize state for humans

The reviewer should not be the same context that drafted the output. Fresh review catches generic phrasing, invented claims, tone problems and missing evidence.

Approval Is a Feature

Approval gates are often treated as friction. They are actually a selling point.

In legal, HR, finance, real estate, healthcare, job search and public-sector-adjacent work, the customer does not want uncontrolled action. They want acceleration without losing control.

The product promise is:

> The agent prepares the work. The human approves the action.

That single sentence removes half the fear.

What To Build

The best agent systems will feel less like magic and more like reliable operations:

• state is readable
• decisions are logged
• proof is attached
• errors are recoverable
• costs are capped
• humans can take over

That is the boring architecture that makes agents useful.