A critique gate tripled agent success on one-third the compute

An agent that fails three tasks out of four is a demo. One that finishes four out of five starts to look like a product. On October 26, 2026 ⁰², a Sungkyunkwan University team reported moving an agent most of that distance.

The team, led by professor Minhae Kwon, reports that success on complex virtual-environment tasks rose from 24.6% ⁰⁶ to 78.3%. Kwon's group says its framework did it on about one-third of the compute. Less compute, far more finished work.

Compare that with Google's Android Bench, which lists GPT-6.1 Sol at a 29.3% ⁰¹ pass rate and a $99.50 ⁰⁵ average cost. A frontier model at frontier prices still misses most of those agentic tasks.

The result comes from an arXiv preprint. Even so, our read is that the wiring around the model has become the cheapest place to buy reliability.

Gated critique spends checks where risk lives

The Sungkyunkwan team calls the framework SAG, short for Self-improving Agent with Gated critique. A critic is a second opinion: a model call that reviews what the agent plans to do. Most agent loops ask for that second opinion on every step. SAG asks only when the answer is likely to change the outcome.

The team describes four moves. The agent estimates when it is uncertain. It requests critique only when feedback should improve the result. It reuses runs that worked. And as that store of experience grows, it leans less on outside feedback.

Those four moves fall into two parts. The gate decides when to spend a check. The memory holds what the agent already knows, so it can skip asking. In practice, checks are rationed, and risk decides who gets one.

The Sungkyunkwan figures put failure at 75.4% ¹⁰ before and 21.7% after, a cut of about 71% ²⁵. By our math from those figures, success per unit of compute rose about 9.5x ²⁶: 78.3 divided by one-third, set against 24.6 divided by one. That multiple holds only if both systems faced the same tasks under the same scoring.

Inside the loop that tripled success

Picture a new warehouse picker on day three. A good one grabs the obvious boxes alone and radios the floor lead only when a bin label looks wrong. A weak one radios on every box, or never radios at all. SAG builds the good picker.

The loop has four steps. First, the agent picks its next action. Second, it scores its own confidence in that action. Third, if confidence is low, it calls the critic before acting. Fourth, it files every successful run in memory for next time.

Say a baseline attempt costs $1 ²⁷ in compute. At the 24.6% ⁰⁶ rate Kwon's team reports for conventional approaches, each finished task costs about $4.07 ²⁸. A gated attempt costs about 33 ³⁰ cents, and at the 78.3% rate the team reports, each finished task costs about 43 cents. The budget is the same, and the finished work is roughly nine and a half times larger.

If Google's $99.50 ⁰⁵ average on Android Bench is a per-task figure, each GPT-6.1 Sol pass at 29.3% ⁰¹ costs about $340 ³². The two tests differ, so treat the comparison as directional. Still, buying reliability through model size alone gets expensive fast.

Why does the gate pay off so well on long tasks? Because small misses compound. Here's an illustration of our own: a 20-step ³³ task where each step succeeds 95% of the time finishes about 36% of the time. Lift each step to 98% ³⁴ and the finish rate climbs to about 67%.

A gate that catches the few risky steps lifts the per-step rate right where the losses pile up, and it skips paying for reviews on the easy steps. That is the whole trick, and it works on whatever model you already pay for.

The gate is also the weak point. Miss a risky step and an error sails through. Flag too many and the savings vanish into review calls. Nobody knows yet whether the 78.3% ⁰⁶ figure survives messy production work, where interfaces change and some actions cannot be undone.

Speed is the other open question. Deciding whether to ask for a review has its own overhead. Wall-clock time may fall more slowly than compute does, and in a live customer chat, seconds can matter as much as cents.

Structure matters too. Google Research evaluated 180 ¹¹ agent configurations and found that independent agents amplified errors by up to 17.2 ¹² times. Centralized systems held amplification to 4.4 ¹³ times, by the same Google Research count. SAG routes every check through one gate, which puts it closer to the centralized pattern.

We think the gate is the most portable idea in the paper. You can bolt it onto almost any agent you already run and measure the result in a spreadsheet.

Reliability is moving into the harness

Agents get bought per success by 2031

Market forecasts are loud, and they disagree. Research and Markets projected on October 21, 2026 ¹⁴ that agentic AI grows from $19.33 billion in 2026 to $205.88 billion ¹⁵ by 2033. Fortune Business Insights puts the 2026 AI-agent market at $11.78 billion ¹⁶. Polaris puts it at $10.2 billion ¹⁷. For a single year, that's a spread of close to 1.9x ³⁵, so read every one of these numbers as a direction, not a measurement.

The stronger signal is what benchmark makers now measure. MLCommons is building MLPerf Agents, an NSF-backed effort funded at $3.5 million ⁰⁹ over two years. It scores the whole agent, including the harness and failure recovery. (A harness is the code that wraps the model, such as the loop and the checks.)

Buyers are already paying for loose wiring. Redwood Software surveyed 300 ¹⁸ senior leaders and practitioners, and 77% ¹⁹ said they had paid a price for ungoverned agents. Precedence Research cites a 2026 survey in which 75% ²⁰ of organizations were piloting or running some AI agent. Only 15% ²¹ were at any stage with fully autonomous ones.

Put the dated pieces side by side and they point the same way. Google's Android Bench lists a frontier model at 29.3% ⁰¹, while Sungkyunkwan's gated agent reports 78.3% ⁰⁶ on its own test. The benchmark buyers will cite is being built to score the system around the model. By 2031, five years on from these results, we expect contracts to price agents per finished task.

The napkin math shows why. On our $1-per-attempt ³⁶ illustration, 10,000 finished tasks a month cost about $40,650 on the baseline. With the gate, the same 10,000 ³⁷ cost about $4,260. Put those two numbers on two vendor quotes and the buyer picks the second one.

The risk is lopsided. A gate costs a weekend of engineering and a few extra calls on hard steps. Skipping it could mean paying close to 9.5x ³¹ more per finished task, if the Sungkyunkwan result holds up. Stronger base models may absorb part of this gain, and Google Research's error data says the design has to fit the task.

Add a critique gate to one agent

Pick one agent you already run and give it a gate over a single weekend. Any loop that calls a model and then acts will take it.

Step one, choose 30 ⁰⁸ tasks the agent already handles, each with a pass check. A pass check is a test a script can grade, such as "the file exists and the totals match." Run all 30 and log pass or fail, plus spend per task.

Step two, add a confidence score: before each action, ask the model to rate from 1 to 5 how sure it is. That rating is a crude uncertainty estimate, meaning the agent's own guess that it is about to slip.

Step three, add the gate: call a critic only when the score is 3 or lower, or when the action cannot be undone. Sending an email or making a payment counts as cannot-be-undone. Run the critic on a different model from the agent, so the two do not share the same blind spots.

Step four, add memory: when a task passes, save its action sequence to a file tagged by task type. On the next run, load the closest saved sequence into the prompt.

Step five, rerun the same 30 ⁰⁸ tasks. Divide total spend by passes to get cost per success, and compare it with step one.

Expect the first run to disappoint, and plan for it. Read every failed task where the gate stayed shut. Those are false negatives, risky steps the gate waved through. Raise the threshold to 4, rerun, and keep a sheet with pass rate, spend, cost per success and critic calls.

Rerun the sheet every week as you add tasks. If cost per success drops and pass rate holds, move the gate to a second agent. If cost climbs, the critic is firing too often, so lower the threshold and test again.