The agent has a plan. Can the user see it?

The most important screen in an AI agent product may not be the chat box.

It may be the plan.

A chat box is good for instruction. It is weak for oversight. Once an AI agent starts doing multi-step work, the user needs to understand more than the prompt and the final answer. They need to see the agent's intended route.

What will it do first? What data will it use? Which systems will it touch? Which steps are reversible? Where will it pause? What will happen if the plan is wrong?

Without that visibility, the user is not delegating. They are hoping.

Agents change the UX problem

Copilot-style AI can often live inside existing interaction patterns. It drafts, summarizes, suggests, classifies, or answers.

Agentic AI changes the product problem because the system does work across time, tools, and consequences. It may inspect records, compare sources, make decisions, prepare outputs, and trigger actions.

This is why Gartner's prediction matters: more than 40% of agentic AI projects will be canceled by the end of 2027 due to escalating costs, unclear business value, or inadequate risk controls. Agentic products do not fail only because the model is weak. They fail because the workflow around the model does not make autonomy usable.

Planning visibility is one of the missing pieces.

A plan is not a progress indicator

Many AI products show activity:

"Thinking..."

"Searching..."

"Generating..."

"Almost done..."

That is not a plan. It is a loading state with better vocabulary.

A plan is specific enough for a user to judge. It describes steps, sources, outputs, decisions, and points of control. It separates work that can run freely from work that needs human approval.

A useful agent plan might show:

• Objective: what the agent is trying to complete.
• Inputs: files, records, instructions, customer data, constraints.
• Steps: the route it intends to take.
• Tool use: which systems it will access.
• Gates: where it will pause for approval.
• Risk flags: where the outcome is consequential or uncertain.
• Preview: what the final action will change.

The plan should be inspectable before execution and comparable after execution.

The user needs to edit the plan

Visibility alone is not enough. A plan the user cannot change is just a nicer warning.

If the agent proposes the wrong route, the user should be able to:

• remove a step
• add a constraint
• change a source
• mark a step as approval-required
• reduce the scope
• ask for a dry run first
• stop the process

That is the difference between passive transparency and active control.

Microsoft's HAX guidelines talk about supporting interaction across phases: initially, during interaction, when the AI is wrong, and over time. Planning visibility sits in the "before" and "during" moments. It gives the user leverage early, when changes are cheap.

Plans create accountability

A visible plan also makes agent work auditable.

If the product stores the intended plan, the executed steps, the deviations, the approval decisions, and the final committed actions, the business can review the system as work rather than mystery.

That matters in serious SaaS contexts. When an AI agent touches customer operations, supply plans, procurement workflows, support queues, financial approvals, or internal communications, the question will not be "was the model clever?" It will be "what exactly happened?"

The plan is the first layer of that answer.

Bad plan visibility creates false confidence

The most dangerous agent is not always the one that makes obvious mistakes. It is the one that appears fluent while hiding the shape of its work.

Users can be impressed by speed and still distrust the system. Or worse, they can trust it until something breaks and then discover there was no useful trace of how the work happened.

Stanford HAI's 2026 AI Index notes that AI agents improved on OSWorld, a benchmark for real computer tasks, but still fail roughly one in three attempts on structured benchmarks. That is not an argument against agents. It is an argument for product designs that assume agents need supervision.

If the agent may fail, the plan has to be visible.

A practical design rule

For every AI agent workflow, ask:

Could a user approve the plan before the agent begins?

If not, the workflow probably lacks planning visibility.

Then ask:

Could a reviewer reconstruct the plan after the agent finishes?

If not, the workflow probably lacks accountability.

The best agent UX supports both. It gives the human a way to steer before work happens and a way to understand what happened afterward.

The WFK position

The future of agent UX is not a better chat box. It is a better operating surface.

Plans, gates, previews, logs, scopes, and recovery paths are the interface of useful autonomy. Chat may start the work, but architecture makes it safe enough to trust.

If the agent has a plan, the product should show it.