Humans build trust by explaining why they made a decision.
What if AI agents briefly summarized their plan before executing?
For example:
“I found three scheduling conflicts. I’ll resolve the first automatically, ask you about the second, and postpone the third until Friday.”
Would this make AI more trustworthy?
Or just slower?
hi @Saiganesh
Yes, I think exposing the plan increases trust, but there is an important distinction. Users dont really need the raw reasoning, they need the plan. Raw chain of thought is verbose and research shows its often not even faithful to what the model actually did, so showing it can create false confidence. A short structured plan like your example is diferent, its verifiable and gives the user a chance to object before anything happens.
The key thing tho is that transparency alone doesnt build trust, control does. A plan you can read but cant veto is just theater. Thats why I’d pair it with human in the loop on the risky actions. In LangChain you get this basically for free, create_agent with HumanInTheLoopMiddleware lets you mark which tools need approval, and the user can approve, edit or reject the proposed action. Deep agents even ship the planning part by default, the write_todos tool externalizes the plan into state and the UI can render it live. Claude Code and Cursor converged on the same todo pattern independently, which tells you something, it also makes the agent itself more reliable, not just more legible.
About the latency question, split it in two. Streaming the plan costs a few dozen tokens, thats noise, and it actually makes the agent feel faster, watching progress beats staring at a spinner for 30 seconds. The only real cost is waiting for the human at an interrupt, but thats the feature, not overhead, and you controll how often you pay it. My rule of thumb, auto execute reads and reversible stuff and just stream what youre doing, interrupt only on irreversible actions like sending emails or spending money. If you interrupt on everything users get approval fatigue and start rubber stamping, and then you lost both speed and oversight.
Here is an example of an existing software system which allows users to approve, edit, or reject proposed plans. In the following example, software developers can approve, edit, or reject plans in a software IDE: Plan Before You Build: Introducing the Plan agent in Visual Studio - Visual Studio Blog . Software developers can manually edit a presented plan (in markdown), refine it through conversation with AI agents, or share it with their team for review.