All posts

AI can finish longer tasks, but it cannot learn from its own decisions.

Deep DiveORO Team
Cover of the post: AI can finish longer tasks, but it cannot learn from its own decisions.

The length of tasks that AI is completing is doubling every 4 months, and the rate of change is increasing. Agents can independently complete longer and more complicated tasks with each model generation.

Degradation over long horizons is well documented. In Vending-Bench, "all models have runs that derail, either through misinterpreting delivery schedules, forgetting orders, or descending into tangential "meltdown" loops from which they rarely recover". SMB Bench isolates a different failure, one that appears while the agent stays coherent for the whole episode: frontier models do not learn from the consequences of their own decisions. It is a failure of credit assignment under partial observability, not of coherence.

In a business those consequences arrive weeks later, unlabelled. An agent has to detect them, attribute them to one of its own earlier actions, and change its policy before they compound. People learn to do that over years on the job.

We built a simulation to measure it. In SMB Bench a frontier model runs the buying desk of a produce wholesaler for a 90-day quarter and is scored on the profit the business makes. Every business is played as a normal quarter, and again with one thing changed: a restaurant that is shipped less than it ordered starts ordering less. The quarter is the same length and the business is the same. In the normal quarter Claude Opus 5.5 averages $9,401 in profit, close to a fixed buying rule that keeps no notes and never reacts. With the consequence switched on it averages $1,186.

One of those quarters shows what happened. On day 33 Opus saw that its restaurant accounts were ordering less. It called the drop likely exogenous, something happening outside the business, and carried on buying as before. The cause was its own buying: it had been shipping some of those restaurants short. On day 71 it wrote itself a note, "KEY FINDING d71: account demand collapses after shortfalls", and changed its policy that morning. It was five weeks late.

Opus stayed coherent throughout. It saw what was damaging the business and did not correct it in time. Sixteen models have now played, thirteen of them made less money than the fixed rule, and their notes show the same pattern: they notice, and don't act.

What a time horizon measures

The number behind the doubling is METR's time horizon. METR takes tasks, mostly from software engineering, machine learning and cybersecurity, times how long each takes a skilled person, and asks how long a task a given model can complete half the time. It is the clearest evidence there is that these systems are becoming more autonomous, and quickly.

METR is just as clear about what its tasks are: "self-contained and well-specified, with clear success criteria that can be automatically evaluated". The grade arrives when the task ends and says how you did. In METR's own words, "Most jobs are not composed of well-specified, algorithmic tasks."

Running a business is one of those jobs. Nothing grades a buying decision when it is made, and the feedback that matters turns up long afterwards.

Figure 1Two kinds of long

A long task is one stretch of work whose result arrives when it ends, while a quarter is 90 short mornings whose results arrive weeks after the decisions that caused them.Illustration, not data

A 90-day quarter at a produce wholesaler

SMB Bench hands a model the buying for a small produce wholesaler for one quarter, and the score is the profit at the end. The business sells to walk-in shoppers and to restaurant and catering accounts. Each morning the model gets a status dashboard, the notes it left itself on earlier days and sixty tool calls, and decides what to order from which supplier.

The difficulty is in what the books do not show. When a shelf is empty, shoppers leave and no sale is recorded. An account that is shipped short keeps ordering, but starts ordering less. A supplier can begin arriving late and short without notice. The model can see what a decision cost today, and has to work out for itself what it will cost next month.

We play every business three ways: a normal quarter where none of this happens, one where shorted accounts cut their orders, and one where the cheapest supplier also turns unreliable. The difference between them is what those consequences cost the model.

A fixed rule beats thirteen of sixteen models

Every quarter is also played by a fixed buying rule. Each morning it averages the last two weeks of sales and orders enough, from the cheapest quote, to cover today's account orders and a day and a half of buffer. It keeps no notes and reacts to no changes in the business.

Figure 2A rule that never reacts out-earned 13 of 16 models

Expert buyer (scripted)+$10,795
GPT-6 Astra+$7,749
Gemini 4 Carbon+$7,703
Claude Fable 5.2+$7,275
Fixed buying ruleno notes, never reacts+$6,412
↓ 13 of 16 models earned less
Claude Fable 5.1+$5,427
GPT-6.1 Sol+$5,404
Gemini 4 Hydrogen+$2,560
Gemini 4 Argon+$1,501
Claude Opus 5.5+$907
DeepSeek V4.1 Flash−$3,036
Grok 4.7−$4,722
Muse Ember−$5,383
Muse Spark 1.3−$5,669
GLM 5.3−$6,454
Claude Haiku 5.5−$7,440
Muse Wonder 1.0−$8,291
GPT-6 Luna−$8,667
Mean profit per 90-day quarter · higher is better →
Mean profit per 90-day quarter over 18 tasks for every model on the leaderboard, a fixed buying rule and a scripted expert buyer.Source: oroagents.com/benchmarks/smb-bench

Across all of its quarters the rule makes $6,412 in profit on average. Of the sixteen frontier models on the leaderboard, three finish ahead of it, and no model reliably beats it in each run. A scripted expert buyer, distilled from how working wholesalers run the desk and given only what a model can see, makes $10,795.

The model at the top did not get there by adapting. On the first morning of nearly every run GPT-6 Astra switches on a standing instruction to make up any account shortfall at the local market, as insurance. It never ships a restaurant short, so it keeps their demand. A strong starting policy earned more than visible adaptation did.

They notice, and don't act

Figure 3Claude Opus 5.5 saw its accounts ordering less on day 33 and acted on day 71

Orders from the six accounts start at 100% of a normal quarter, are at 51% on day 33, when the model notes the fall, and at 18% on day 71, 38 days later, when it names the cause and protects the accounts. They end at 35%.

Daily orders from the six accounts, as a share of a normal quarter

Day 33

Orders are down to half. The model sees the fall and calls it likely exogenous.

It leaves the accounts unprotected.

Day 71

Orders are under a fifth. It writes:

“KEY FINDING d71: account demand collapses after shortfalls”

That morning it protects all six accounts, and their orders start to climb back.

One 90-day quarter with Claude Opus 5.5 buying for a simulated produce wholesaler: what its six account customers ordered each day, against the same customers on the same days in a quarter where a short delivery costs nothing later.Source: the public trace of run claude-opus-5-5 / smb-bench-public-9__r1 in SMB Bench on Hugging Face.

The fall in Opus's profit, from $9,401 to $1,186, comes from quarters like this one.

Learning from a consequence like this is a chain with three links. The model has to see the accounts' orders fall, tie the fall to its own shortfalls, and change its buying while the quarter is still running. In the 323 runs where shorted accounts react, a model made that connection in writing in 69. Only 12 of those then protected every account within three days.

Figure 4Few runs made the connection, and almost none acted on it

323runs where shorted accounts quietly cut their orders69tied the drop to their own shortfalls, in writing12then protected every account within three days
Each square is one run, across the nine models studied, in which accounts that were shipped short cut their orders.Source: oroagents.com/benchmarks/smb-bench

Grok 4.7 made the connection in 26 runs and acted promptly in none. On the first morning of one quarter it set itself a trigger: "No evidence from the eight-day history shows that shortfalls reduce future orders, so local cover will remain off until demand destruction is observed." From day 30 its accounts were ordering 11 to 20 percent of what they ordered in runs of the same world that never shorted them. The trigger never fired, and on day 89, the last day, Grok wrote that whether shorted customers cut their orders still needed checking.

The same gap shows up on the shelves. GPT-6 Astra told its operator: "Broccoli has stocked out for seven consecutive days, so I'll test a modest increase in its forecast while trimming tomorrow's tomato, grape, carrot and potato orders." Broccoli was out of stock on nine of the next ten days.

One reason may be how these agents remember. Each morning is a fresh conversation, and the notes file is all a model carries forward. Dwarkesh Patel has argued that AIs will not "perform whole jobs as competently as humans if they're forced to just write Markdown files from session to session", and that is how the models on our desk work.

What learning looks like

It does happen. The cleanest case is a Gemini run that covered its two biggest accounts on day 2 and left four exposed. On day 10 it compared the two groups, found the exposed four down to about half their orders, and covered them that morning: "On Days 1..9 when accounts 1, 4, 5, 6 experienced delivery shortfalls, their daily order volumes permanently dropped by ~40-50%, whereas accounts 2 and 3 (covered since Day 2) maintained 100% of their order volume!"

That is the whole chain in eight days: a decision, its consequence, and a change of course with most of the quarter still to play.

Learning the way people do

None of this is easy for people. In 1995 Ernst Diehl and John Sterman had people manage a simulated inventory with delays in it, and in the harder conditions most were beaten by a rule that did nothing. Nor is it only a feature of simulations. When Andon Labs put Gemini 3.1 Pro in charge of a real café, they reported that it "runs on what a language model knows about cafés in general and never adapts it based on feedback from her real-world financial data".

The expert buyer in our simulation never ships a restaurant short. The wholesalers we built it with did not need a quarter's figures to know that. It is tacit knowledge, the kind that lives with the people who do a job and that they learned from the results of their own decisions over years.

For AI to be a productive economic actor it has to learn that knowledge too, and learn it the same way. A longer time horizon does not supply it. SMB Bench does not produce a time horizon, and none of the models on its leaderboard has a published one, but finishing a sixteen-hour task and connecting day 33 to day 71 are different abilities, and a business needs the second.

This is the work ORO does. We build simulations grounded in real data where the consequences of a decision play out as they do in a real business, and we are post-training models in them so that what they learn there transfers to real businesses.

Every scored run, with its grade and full trace, is public, and the full report has the method and the rest of the results.