An agent's capability comes from three places: the knowledge and reasoning the model acquired in training, the context the task provides, and the tools it can call. It fails in four situations. The longer the task, the more small errors at each step accumulate. When a task lies outside the boundary of its capability, the agent appears just as confident. When it lacks a basis, it tends to give a plausible-looking answer rather than say it does not know. When there is too much information, the key content gets overlooked. Reliability determines how much time the principal must spend finding errors and reworking them. The way to raise reliability is to shorten every stretch of the process that runs unchecked, have the agent stop when it is not confident, and use tools to verify whatever can be verified.
A task such as "compare the fees and maximum drawdowns of two funds over the past three years" needs three things. The model provides the ability to understand the question, organize the steps and write the conclusion. The context provides the information the task needs, such as the funds' names and the horizon the principal cares about. Tools provide what the model cannot do by itself, such as querying the latest net asset values and calculating drawdowns precisely. If any of the three is missing, the result goes wrong: without context, the answer can only be general; without tools, the numbers may be made up from memory; with an inadequate model, the steps themselves go wrong. Research by Schick et al. (2023) and Gao et al. (2023) shows that letting a model call a calculator, run a search or execute a program clearly reduces calculation and factual errors: tools both extend capability and provide a means of verification.
Three sources of capability
Task: compare the fees and maximum drawdowns of two funds over the past three years. Tick or untick the three sources and see how the agent's result changes.
A task consists of many steps, and each step has some probability of error. If each step is 99% correct, the chance that a 10-step task is entirely correct is about 90%; for a 100-step task it is only about 37%. This is why long tasks are hard to hand to an agent: reliability per step looks high, but accumulated it is low. Adding checkpoints to the process changes this: verify after each segment and correct errors as soon as they are found, and errors are not passed all the way down. The more easily checks can be automated, for example by checking calculations with a program or code with tests, the more reliable long tasks become.
Every step is reliable, but the whole may not be
Adjust the accuracy of each step and the number of steps, and see the probability that the whole task is entirely correct. Turn on checkpoints: verify every 10 steps, and 90% of errors are found and corrected.
Researchers at METR measure an agent's capability by its "time horizon": the time a human professional would need for the tasks the agent can complete with a given level of confidence (Kwa et al., 2025). They found that the longer a task takes a human, the lower the agent's success rate. Between 2019 and 2025, the length of tasks agents could complete with 50% confidence doubled roughly every seven months; since 2023 the pace has been faster, doubling about every four months, and the measurements for the first half of 2026 are clearly above the level extrapolated at seven months (METR, 2026). For the same agent, the tasks it can complete with 50% confidence are several times longer than those it can complete with 80% confidence. For the principal, the latter figure is the more meaningful one: a task with only even odds of completion still has to be checked item by item after it is handed over.
The boundary of an agent's capability is not even. Dell'Acqua et al. (2023) had 758 consultants at Boston Consulting Group complete a set of tasks. On tasks within the agent's capability, consultants using the agent completed more, faster and at higher quality. On a task that looked similarly difficult but lay outside that capability, consultants using the agent reached the correct conclusion about 19 percentage points less often than those not using it. They call this the "jagged technological frontier": which tasks are inside the boundary is hard to judge from how a task looks, and the agent shows almost the same confidence inside and outside it.
When an agent lacks a basis, it often gives a fluent but wrong answer; this is called hallucination. Kalai et al. (2025) explain its origin in scoring rules. Most evaluations score only the share of correct answers: answering "I don't know" scores zero, and so does a wrong answer. When unsure, guessing therefore always scores higher than admitting ignorance, and a model trained on such evaluations learns to guess. If wrong answers are penalized and admitting ignorance is not, answering is worthwhile only when confidence is high enough. For financial products the rule should be explicit: with no data, say there is no data; when something cannot be determined, say where the uncertainty lies.
Scoring rules decide whether an agent guesses
The agent has some confidence about a question. Choose a scoring rule, drag the level of confidence, and compare the expected score of answering with that of saying "I don't know".
Giving an agent more context is not always better. Liu et al. (2024) found experimentally that a model uses key information well when it sits at the beginning or end of a long series of material, and noticeably worse when it sits in the middle, a U-shaped pattern. The longer the context, the stronger the effect. In a financial task, if an annual report, dozens of news items and a year of trading records all go into the context, the one item that matters may be buried. The more reliable approach is to first select the information relevant to the task, then hand it to the model, and require the model to mark which piece of material each conclusion comes from.
Key information in the middle is the easiest to overlook
The model is given 20 documents, only one of which contains the answer. Drag the position of that document and see the accuracy of the answer; turn on "Select relevant documents first" to compare.
Reliability comes from the design of the process, not only from a stronger model: shorten the stretches that run unchecked, have the agent stop when it is not confident, and use tools to verify what can be verified.
Numbers come from data sources and calculation, with source and cut-off time noted; they are not supplied by the model from memory.
Break long tasks into segments that can be verified, checked automatically or by the user at the end of each.
Say plainly which tasks the agent is known to handle poorly, and hand them to a person or switch to more conservative handling.
With no data, say there is no data, and mark uncertain points separately; penalize wrong answers in evaluation.