Agents need control planes, not just prompts

May 5, 2026

The first interface to an agent is usually a prompt. That is fine for the first demo, but it is not enough for a real product.

Once an agent starts doing meaningful work, the useful object is no longer a single message and a single answer. It is a session: something that can be started, resumed, interrupted, inspected, steered, rolled back, and reviewed. The agent becomes less like a text box and more like a runtime.

That changes what has to be designed. A product needs to know what the agent is doing right now, which tools it called, what it changed, whether it is waiting for approval, how to recover after an error, and how to show progress without creating noise. These are not prompt-engineering problems. They are control-plane problems.

The same pattern shows up in coding agents, personal agents, and workplace agents. At small scale, chat is enough. At larger scale, people want lifecycle: list sessions, resume a thread, fork a direction, interrupt a bad run, review a diff, approve a command, archive a completed task, and understand what happened later.

This is why protocols and harnesses matter. The surrounding system determines whether the agent can survive real workflows. The model may be powerful, but the product only becomes trustworthy when state, events, tools, permissions, and recovery paths are explicit.

Reliability also feels different in this world. A lifecycle bug is not just a UI glitch. A stuck run, duplicated job, missing transcript, orphaned progress update, or dropped final delivery can become a control failure. The user does not care that the model produced the right internal artifact if the final handoff failed.

My current bias is that many strong agent products will look less like smarter chat windows and more like well-designed control surfaces. The chat may still be there, but underneath it there should be a real session protocol: events, tools, approvals, filesystem or workspace access, streaming progress, and a way to recover when something goes wrong.

Prompts are how we talk to agents. Control planes are how we live with them.