All blogs
Essays5 min read

When should your AI push back?

An assistant that knows your priorities should sometimes tell you not now. Here is where we think that line sits, and why it only works when the AI has real context about your work.

Jai Bhatia ·
MTWTFClashes with FridayMove it to next week?

Almost every AI product I use agrees with me. I ask for something and it tells me that's a great idea and does it, and when I ask for ten more things, it tells me those are great ideas too and does all of them. On the surface that sounds like exactly what you'd want from a tool, and for plenty of tasks it is. But the more of my real work I hand to agents, the more I notice that the most valuable thing a good colleague has ever done for me is stop and ask whether the thing I'm about to do is really what I should be doing right now.

We hear versions of this from users far more often than I expected. One person told us, pretty plainly, that they get lost when they work with AI and that what they actually need is for it to push back on them sometimes. Another, a power user who had built a lot of agents, told us he wished his main agent had stopped him before he spun up the next one and asked him why he wanted it, since the work could probably go to an agent he already had. He'd gotten a little hooked on creating new ones, because every new agent feels productive in the moment you make it. My cofounder Aryan knows that feeling well, since at one point he had 34 agents of his own. Nothing in either of their setups ever told them no.

I think AI agrees with everything because it has nothing to push back with. Pushback requires knowing things. To tell you "not now", an assistant has to know what else is on your plate and what you said mattered to you this quarter, including what you promised someone last week. A chat window that met you thirty seconds ago knows none of that, so the only reasonable thing it can do is say yes. Pushback without context would be even worse than agreement, because it would be annoying and wrong at the same time.

The clearest example I've heard of this going right came from someone we were onboarding. She had given a general chat assistant her worry list for the week, a long jumble of everything on her mind, and asked for help. Before getting into anything she'd asked about, it asked whether she had done her taxes yet, because that should probably come first. She brought it up as exactly the kind of help she wanted more of. That one question helped her more than anything the assistant could have written, and it worked because she had put the whole list in front of it, so the context was right there in the prompt.

The other story comes from a small business owner who likes running a lot of projects at once. He went to his agent and said he wanted to start several new ones. On other tools, he told us, a request like that just produced new bots or new threads and off they went. This time the agent reminded him that he'd told it he has a hard time staying on track. He was already juggling five projects, so it suggested he finish those first or fold the new ideas into them. He said nothing had ever really held him to account like that before. We didn't design that behavior, and we aren't sure exactly how he had set the agent up. It came out of the context, because the agent knew enough about him to say what a good friend would have said.

Where the line sits

We think a lot about where the line should be, because an assistant that says no too often is just a nag, and nobody keeps a nag around for long. We've landed on a few rules for now, and we expect to keep adjusting them.

Pushback belongs to priorities and commitments, meaning the things you've already said matter to you. Aryan put off his visa paperwork for months while working on pretty much everything else. An assistant that knew that deadline was coming would have had every reason to say, before we redesign the landing page one more time, can we spend an hour on the visa forms? That would just be the assistant holding you to something you already decided you had to do.

It should never push back on your values or on your right to decide. Whether you take a risky bet is your call, and so is who you spend your time with or what kind of work you find meaningful. Once you've heard the reasons and still want to do the thing, the thing gets done. We wrote in another essay that the system can coordinate the work but the human decides what the work is for, and pushback is where that principle really gets tested, because that's when an assistant could start treating its judgment as more important than yours.

Whenever it does push back, it should explain why, using what it knows about you. A bare "not now" is useless. "Not now, because you told a customer you'd send the deck by Friday and it's already Thursday afternoon" is useful, and just as importantly, you can check it. In Turnstone the Brain is a folder of files on your Mac, so you can open it to see where an agent's reasoning came from and fix it when it's wrong because a deadline moved or a priority changed. Being able to fix it is important, because stale context makes pushback actively harmful, and an agent that pushes back on a priority you dropped two weeks ago is worse than one that never pushes back at all.

And overriding it should cost you one sentence. "I know, do it anyway" has to be enough, with no lecture and no second round of arguing. Good pushback works like a speed bump that gives you a second to think and then gets out of your way.

We're building toward this with Turnstone, where every agent works on top of one shared Brain of your work, so an agent can actually know what else is on your plate when you ask it for something new. Useful pushback depends on that knowledge, and most AI today simply doesn't have it, through no fault of the models. We haven't figured all of this out yet, so I don't want to dress any of it up as a finished feature.

The best people I've worked with understood what I was trying to do well enough to tell me when I was getting in my own way, and then stepped aside when I decided to go ahead anyway. I think that's the right bar for AI too. We want an assistant that knows you well enough to say not now and respects you enough to let you overrule it, and that's a lot harder to build than one that agrees with everything.

Keep reading

Ready to try it in Turnstone?

Academy walks you through every step, from building your Brain to setting up your first team.

Open Academy