Why assistants should ask instead of guessing
Giving an AI assistant more tools doesn't make it more competent. Assistants do better when they know where the decision points in a task are and ask a focused question at those points, instead of guessing and trying tool after tool.
A common response to an assistant that gets stuck is to connect more tools and APIs. An assistant with fifty tools and no sense of when to stop and check tends to try them one after another, making changes nobody asked for. A short, well-timed question to the person it is working for often settles in seconds what trial and error can't.
The overloaded toolbelt
In the world of AI assistants, software makers are currently repeating the mistake of the clattering apprentice. Whenever an assistant struggles to finish a long errand or solve a difficult coding problem, the standard response has been to equip it with more software tools.
We give the assistant database connectors, search endpoints, spreadsheet parsers, shell runners, and documentation libraries. Before the assistant even reads your first request, its memory is stuffed with dozens of tool descriptions, parameter lists, and formatting rules.
That approach sounds generous, but it comes with a steep hidden cost. When a model has to choose among dozens of overlapping tool descriptions, it picks the wrong one more often. The problem is competition between overlapping choices more than the total length of the text.2
Faced with forty plausible buttons to press, the assistant begins guessing. It runs one query, gets an unexpected error, tries a different API, reformats the output, and wanders into a loop of blind trial and error. By the time it stops, it has burned through minutes of computing time and produced a messy workaround that nobody can safely audit.
Two ways to tackle an open road
The power of the fork in the road
What if software did not start by reaching for a wrench? What if it started by identifying the forks in the road?
In a recent benchmark of complex engineering tasks, researchers compared two different setups.1 In the first setup, an autonomous agent received a large suite of performance tools, parsers, and system documentation. In the second setup, called KernelSeek, the agent was given almost no specialized external tools. Instead, it received a simple bank of structured questions at key decision points: Where is the bottleneck? Does this memory layout align with the hardware cache? What tradeoff are we making?1
The results were striking. The tool-heavy agent quickly plateaued, trapped in loops of guessing and tweaking. The question-driven agent easily surpassed it, discovering faster and cleaner solutions without needing pre-built recipes or bulky software libraries.1
The lesson applies far beyond writing computer code. When you ask an assistant to plan a family trip, manage clinic follow-ups, or balance your business books, the hardest part is rarely clicking the final submit button. The hard part is deciding which path to take when two good options point in different directions.
Three moments where asking beats guessing
To see how this works in practice, look at three everyday situations where an assistant ought to stop and ask.
1. The clinic appointment with an incomplete record. Imagine a community clinic using an assistant to schedule follow-up care for an elderly patient. The doctor's referral note mentions two possible clinics: one near the patient's daughter and one near their bus route. A tool-stuffed assistant might guess based on the latest calendar entry, booking a slot without checking whether the patient has a ride on Tuesdays.
A question-driven assistant does not guess. It notes the ambiguity, drafts a short message to the clinic coordinator, and asks: "The patient has two preferred locations on file. Which one should we prioritize for the appointment?" One quick human answer prevents a missed appointment and a wasted afternoon.
2. The small business supplier quote. Suppose you manage inventory for a local bookstore and use an assistant to order next month's reading group picks. The distributor's automated portal offers a 15% discount if you order twenty extra copies, but the return policy is non-refundable.
An assistant loaded with payment tools might simply see the lower unit price and authorize the larger order. A disciplined assistant recognizes a financial risk fork. It pauses, summarizes the tradeoff in one sentence, and asks you whether your cash flow can handle twenty unsold copies if the group changes its mind.
3. The household errand with hidden tradeoffs. Picture asking an assistant to hire a technician to repair a leaking washing machine before the weekend. The assistant finds an available technician who charges double for Saturday morning calls, and another who charges standard rates on Monday afternoon.
Instead of guessing whether you would rather save money or get clean laundry sooner, the assistant sends a two-choice poll to your phone. You tap your answer while standing in line at the grocery store. The errand moves forward without surprises.
What this means
A helpful assistant does not pretend to read your mind. By asking the right question at the right moment, it saves you time while leaving every important judgment in your hands.
What to ask before you buy or build
If you are looking at an AI assistant for your team or household, you do not need to inspect its source code to know if it is well-designed. You can ask four simple questions:
Does it have a way to say "I don't know"? If an assistant always produces an answer, even when given incomplete or conflicting facts, it is guessing. A trustworthy assistant has a built-in brake pedal.
Can it ask you a question before spending money or changing a record? Every action that changes state (charging a card, deleting a file, or sending a formal notice) should require an explicit waypoint. We explore how scoped permissions work in Who is actually acting? and AI agents should hold permissions that expire.
Does it show its reasoning receipt? When the assistant asks you a question, does it show which two options it considered and why? You should always be able to inspect the question waypoint.
Can you turn off tools it does not need? If an assistant only needs to read appointment dates, it should never have access to your email contacts or master payment login.
Helpful technology should make our lives simpler, not noisier. By replacing bloated toolbelts with thoughtful questions, we can build assistants that feel less like clattering machinery and more like reliable human partners.
1. Gabriele Oliaro, Jaeseong Lee, Zhaoyuan Su, Sreeram Vennam, Zhihao Jia, and Samyam Rajbhandari, Ask, Don't Equip: Question-Driven Coding Agents for Deep Kernel Optimization, arXiv:2611.03517 (November 2026).
2. Yinuo Liu, Length Is Cheap, Competition Is Not: Attention Predicts Context Rot in Retrieval, arXiv:2611.04169 (November 2026).
