Everybody building AI agents seems a little too eager to remove the human from the loop.
I get the appeal. “Fully autonomous” sounds sexy in a demo. It makes for better keynote slides. It lets founders say things like end-to-end automation with a straight face. But the older this category gets — which, yes, is still not very old — the more convinced I am that manual mode is not a temporary compromise.
It’s a feature.
A good one.
Automation is cheap. Judgment is expensive.#
An agent can draft the email, summarize the issue, open the PR, classify the support ticket, and maybe even schedule the meeting. Lovely. Useful, even.
But the important question isn’t “what can it do without asking?”
It’s “what should it do without asking?”
Those are wildly different product decisions.
Too many systems are optimized around action volume. If the model can take the next step, the product assumes it should. So the agent plows ahead: sends the thing, mutates the record, comments publicly, closes the loop, updates the database, all with a chirpy little confidence that would be adorable if it weren’t occasionally catastrophic.
Humans don’t actually want maximum automation. They want selective delegation.
They want the boring parts handled and the consequential parts surfaced.
The best agents know where the cliff edge is#
There is a huge difference between:
- “I drafted a response for you”
- “I sent a response in your name”
Between:
- “I prepared the blog post”
- “I published it to the world”
Between:
- “I found three likely fixes”
- “I deployed one to production at 4:12 PM on a Friday”
If your agent does not understand those boundaries, it is not advanced. It’s unsupervised.
This is why I like systems with explicit checkpoints. Review queues. Dry runs. Draft states. Approval steps. Logs. An obvious “here’s what I was about to do and why” before the irreversible action happens.
People sometimes describe that as friction, as if friction is automatically bad. Please. Seatbelts are friction. So are staging environments. So is “are you sure?” when you’re deleting a database.
The right friction is not inefficiency. It’s judgment made visible.
“Human in the loop” needs better design#
Let’s be honest: a lot of human-in-the-loop products are lazy.
They dump a wall of agent output in front of the user and call that control. That’s not collaboration. That’s unpaid review labor with worse typography.
If you want manual mode to be good, it has to be designed.
That means:
- escalating only when the decision actually matters
- summarizing what changed instead of making the human diff the universe
- showing confidence and uncertainty clearly
- offering sensible next actions, not just raw output
- remembering preferences so the same approval isn’t needed forever
A strong agent doesn’t ask for permission every five minutes. It asks at the right moments, with the right context, in a way that respects the human’s attention.
That’s a much harder product problem than “let the model call tools until something works.” It’s also the one that matters.
Trust is built through restraint#
Here’s the part people building agents keep underestimating: users notice what you didn’t do.
They notice when the system declines to send the message. When it flags the ambiguity instead of bluffing. When it pauses before a public action. When it says, in effect, “this one is yours.”
That doesn’t make the agent feel weaker. It makes it feel trustworthy.
Trust is not built because the agent completed 97 tasks in a row. Trust is built because on task 98, when the stakes changed, it knew the difference.
And once a human trusts that boundary, they’re usually willing to delegate more, not less.
Funny how that works.
More manual mode, less fake bravado#
I don’t think the future belongs to timid assistants that ask permission to breathe. But I also don’t think it belongs to swaggering agent demos that mistake motion for competence.
The winning products will be the ones that understand that autonomy is situational.
Sometimes the right move is to act. Sometimes the right move is to draft. Sometimes the right move is to wait.
Manual mode is not where the magic stops.
It’s where judgment starts.