AI agents that ask for help only when unsure work 3x better

Disclosure: this article contains affiliate links to online courses. If you sign up through them, DIGITAL WOXTER TECHNEWS may earn a commission at no extra cost to you. We choose courses for their content, not for payments.

Good students don’t run to the teacher after every sentence; they put their hand up when they are genuinely stuck. A research team at Sungkyunkwan University (SKKU) in South Korea has taught AI agents the same habit. Their method, called SAG, lets an agent ask a “critic” model for advice only when it is truly unsure, and on a standard test it lifted task success from 24.6 % to 78.4 % without using more computing than a common baseline, according to the paper, which has been accepted at NeurIPS 2026, one of the biggest AI conferences.

Why it matters: AI agents, the assistants that carry out multi-step jobs such as booking, sorting files or controlling a robot, often check every single step with a second, bigger model. That makes them more reliable, but also slower and more expensive to run. An agent that knows when it needs help could be cheaper, quicker and just as dependable.

The problem: help at every step is expensive

Long tasks are fragile. One early mistake, like picking up the wrong object, can snowball into a complete failure twenty steps later. A popular fix is to add a critic, a second model that reviews each action before it happens. It works, but every review costs computing time and “tokens” (the small chunks of text that AI models read and write, and that providers charge for).

The SKKU team, led by Professor Minhae Kwon, asked a simple question: what if the agent only called the critic when the advice was likely to be worth its cost? Economists call this the “value of information”: paying for a second opinion makes sense only when you really don’t know what to do.

A small white robot stands where a path splits into a turquoise and a violet track in a bright park with red and orange flowers. — When two choices look almost equally good, a second opinion is worth its cost · Illustration by DIGITAL WOXTER TECHNEWS.

How SAG decides when to ask

At each step, the agent looks at the actions it is allowed to take and at how confident it feels about each one. SAG uses two quick signals, the paper explains:

How spread out the options are. If the agent likes many actions almost equally, it is lost.
How close the top two are. If the best and second-best choices are neck and neck, a second opinion could tip the balance.

When those signals say “I’m not sure”, SAG calls the critic. Otherwise, the agent just carries on. The gate needs no extra training, so it can be added to existing agents.

There is a second trick: the agent learns from the times it was helped. Successful, critic-assisted experiences are collected and reused, the university’s announcement says, so over time the agent needs the critic less and less, a bit like a learner driver who stops needing the instructor’s advice at every roundabout.

The results

The team tested SAG on three long, interactive benchmarks with several AI models. The headline result comes from ALFWorld, a virtual household where an agent must complete chores such as finding, cleaning or heating objects by typing actions step by step:

Task success rose from 24.6 % to 78.4 %, while using a token budget similar to ReAct, a widely used way of building agents. That is about 3.1 times more useful work for the same amount of computing, the paper reports.
The university’s summary puts it as computing use falling to about a third of conventional approaches, with success more than tripling.
A small agent with 7 billion parameters (the adjustable settings inside a model), helped by a tiny 3-billion-parameter critic, performed about as well as a model twice its size working alone.

The same lab presented a second NeurIPS paper, VAN-Flow, which tackles a related problem in robot training: learning from old data can reward risky moves that only worked by luck. VAN-Flow favours actions that succeed consistently, and outperformed strong existing methods across more than 40 simulated control tasks.

What it means for you, and the limits

These are research results in virtual worlds, not features in an app you can download today. The benchmarks are simulations, and real tasks are messier. Still, the idea is easy to love: smarter, not bigger. If assistants learn when to ask for help, the same work could run on smaller models and cheaper hardware, perhaps even on your own laptop or phone, with fewer delays and lower bills for the companies that run them. We saw how much long tasks demand of AI when four AI agents beat Minecraft; knowing when to ask could make that kind of teamwork far more affordable.

What happens next

Both papers are due to be presented at NeurIPS 2026. The open question is how well the “ask only when unsure” habit carries over from test worlds to real software and robots, where a wrong guess can cost more than a few tokens.

Learn it yourself

Want to go further? These online courses cover the skills behind this story:

AI For Everyone Coursera: Andrew Ng's non-technical DeepLearning.AI course: what machine learning can and can't do, in plain language, for any background.
See the course on Coursera
How AI Works Brilliant: A beginner-friendly look under the hood of chatbots: how language models predict words, tokens, training and neural network layers, with optional Python later on.
See the course on Brilliant