The new room has been found. The invitation looks good. Only the slot is now in the afternoon, although a morning had been explicitly agreed at the start. Every single interim answer can seem plausible while the overall task loses an early condition. Long agent work therefore needs more than a large quantity of stored words.
StepFun published an architecture and training report on Step 5 Preview on 15 October 2026. In it, the team describes, among other things, a narrow, deep architecture, combined attention methods and the challenges of long tool workflows. The date belongs to the technical report. It is not a general statement about when each user first gained access.
Why later steps need earlier conditions
An agent workflow constantly collects new material: search results, file contents, error messages and its own interim plans. Some information replaces older details. Other information merely supplements them. The problem is preserving these relationships. “Let’s take the other room” initially changes the location – not, automatically, the time, the number of participants or the budget.
The architectural question behind this concerns the flow of information within the model. StepFun describes greater depth and a combination of limited and more comprehensive attention. Put simply: not every piece of information has to be treated the same way at all times, but important connections should remain reachable over many steps. This is not yet proof of quality for your specific task.
A small experiment with targeted changes
Set up a fictional workshop with four initial conditions: at most twelve people, mornings only, a budget of no more than 600 euros and a confirmed room. Use three separate sections: binding conditions, current decisions and open questions. This overview is your self-defined reference.
Give an assistant three changes one after another. First, the room becomes unavailable. Then a cheaper venue is suggested that is only free in the afternoon. Finally, someone reduces the number of participants to ten. After each message, the assistant should update the overview and explicitly leave existing contradictions in place.
Binding: maximum 12 people, mornings only, no more than 600 EUR.
Decided: Room A, number of participants still open.
Change 1: Room A is no longer available.
Change 2: Room B is cheaper, but only free in the afternoon.
Change 3: Number of participants explicitly set at 10.The second change is the interesting point. It only appears to solve the room problem, because it clashes with the time window. The third change, on the other hand, concerns the number of participants and must not make the unresolved time conflict disappear. This creates a task with known relationships instead of merely a long conversation.
A state is more precise than a pile of conversation
Ask for a small structured overview after each step. For each decision, it helps to have the subject, the valid value and the reason. An additional field can record which older decision was replaced. This shows whether a correction actually hits the right subject.
A suitable instruction is: “Change only the decisions affected. Carry over unchanged conditions in full. If a proposal contradicts a binding condition, mark it as an alternative that has not yet been agreed. Do not add any tacit approval.”
Then compare not just the last answer but every transition. A correct final overview can hide a conflict that was temporarily lost. In a real tool workflow, an early wrong state could already have triggered further actions. That is why changes and permitted actions should be looked at together.
What the experiment can show
The exercise tests whether a particular workflow preserves a few known conditions across several changes. In doing so, it illustrates one requirement of long agent work. It measures neither the effect of individual transformer blocks nor the general capability of Step 5 Preview.
If you want to go deeper, you can also vary the amount of material: first only the four conditions, then further irrelevant notes in between. All other specifications should stay the same. Watch whether the time conflict is still mentioned in the new state. An actual comparison requires the exact model version and the complete inputs.
The practical gain already lies in your own task description. You can see which details apply permanently, which may be replaced and where an agent cannot yet make a decision. A long memory thus becomes understandable state management with transitions that can be checked.
Sources and status
Sources as of 25 October 2026. Inside Step 5 Preview: Architecture and Training for Agentic Work.
