On stage Jimmy places 2 numbers side by side and asks one question straight, who would you hire. The first number is the coding exam score, around 96%. The second number is long horizon computer use, sustained work using a computer on its own, around 20.6% only. He leaves just one line next to this pair of numbers.
Good at exams is not the same as finishing real work.
The bottleneck for agents in 2026 is no longer building, it has moved to maintaining. This pair of numbers is where that shift shows up most clearly. What it points to is not just which task the agent is strong or weak at, it is a full redefinition of what measuring even means.
Measured by Time Now, Not by Correctness
Benchmarks used to measure whether the answer was right. Long horizon work changes the unit to time. The metric becomes how long a task keeps running before it fails. The point he cites is where half of all attempts still succeed. Put simply, the question is no longer what percentage the agent gets right, it is how far the agent survives before it starts drifting off course when left to work alone for a long stretch.
This is exactly why 96% and 20.6% cannot even be compared directly. The 2 numbers measure different dimensions. The first measures correctness on a task that has a clean end. The second measures endurance on work with no clear finish line. And endurance is what most real work inside an organization actually needs, more than exam style correctness.
Context Is the Memory Bottleneck Nobody Has Solved
One reason agents fail midway is that context input per call runs in the tens of thousands to hundreds of thousands of tokens, while output sits at only a few hundred tokens. Most of the processing power goes into rereading what already exists, not thinking through anything new. The longer the context, the more working memory it eats. On a large model, 1 long context set eats tens of gigabytes of working memory, and he states plainly that nobody has actually solved this yet.
This explains why long tasks break down midway. It is not that the model forgets what the task is, it is that the longer it works, the heavier the load of carrying everything it already holds becomes, until it reaches a point where it starts drifting off course on its own, without having been wrong from the start.
Why Code Goes Far, Why Screens Move Slow
The reason these 2 numbers sit so far apart is not a coincidence. Coding work goes furthest because it can be verified. The scope is clear, the result is measured directly, and the reward signal follows the same clarity. Run the test and you know instantly whether it passes or fails, no argument needed.
Computer use work lags behind for exactly the opposite reason. The scope is open ended, with no self contained test to run. It has to look at the screen and interpret on its own what is right and what is wrong, and it still runs many times slower than a person doing it directly. Even if the agent never drifts off course once, speed alone is still an obstacle.
Put another way, code looks smarter than screen work not because the model is better at coding than anything else, it is because code offers a way to verify itself that screen work does not.
Scale Up, and the Problem That Shows Up Is Not Capability
Once you move from a single task to enterprise scale, the problem that shows up changes shape entirely. It is no longer about whether the model is good enough, it is about who is actually governing it. Gartner expects the number of agents inside Fortune 500 companies to grow from fewer than 15 to more than 150,000 by 2028. The cost per workflow rises along with it, up to 5 times within that same year, while only 13% of organizations have a governance system actually ready to use. The count grows by tens of thousands of times within a few years, but only 13% can actually govern it. The gap between these 2 numbers is more worrying than the cost figure itself. He sums it up right here, the bottleneck has moved from building to maintaining.
Small Models Become a Market of Their Own
Another signal he breaks out into its own market is small models. Models in the low billion parameter range that run on device without depending on the cloud, what the field calls edge work, get all three benefits in one deal, privacy, low latency, and lower cost. He also says that this small model market is projected to grow at an average of 28.7% per year. This is a forecast, not something already measured, so it should be read as a broad direction rather than a fixed figure.
This connects directly to the governance point. If an organization is already struggling to oversee a large number of agents on the cloud, having some work run on device without sending data out anywhere at all shrinks part of the scope that needs governing. It is not a full fix, but it does reduce risk in certain spots.
The Closing Line That Governs This Whole Piece
The closing line of this section is the most usable line in the whole piece. Agents in 2026 are not short on intelligence, they are short on endurance and short on governance. The next battleground sits on three fronts at once, task longevity, governance, and work running on the edge.
The last post covered picking a framework and always starting with a single agent. This one explains why starting small actually matters. It is not excessive caution, it is that the 20.6% figure states plainly that long tasks are still where agents fail most often, and the longer the task, the closer a human needs to stay watching, not walk away once it starts.
How to Use This for Next Year's Planning
If you are planning to bring agents into your team next year, the first question should not be which model scores highest on the exam. An exam score only tells you how far a task with clear boundaries can go. It does not tell you how far a task left to run on its own for a long stretch will survive.
The work to hand to an agent first is work with clearly verifiable results, like coding. Open ended work like full screen use should wait until there is a checkpoint in place for a human to catch what fails.
As for governance, do not wait until the number of agents on the team outgrows what you can manage before you start putting a system in place. 13% is the share of organizations with a system ready right now. If your team still has no checkpoint for a human to review what the agent does, that means you are in the large majority that is not ready yet, not an exception.
Source
Everything here comes from AI Update Bangkok 2026, held on 21 August 2026 at The Cloud Bangkok by Jimmy Tejasen. This part covers the session on agent endurance and governance.