Answer
When should a long-running AI task be split into multiple sessions?
At a phase change, at a checkable output, and whenever the answers stop being specific. Not at a token count.
At a phase change, at any point where a checkable output exists, and whenever the answers stop referring to your specifics. Splitting at a capacity limit means the boundary was chosen by the tool rather than by the work.
There are two ways a session ends: because the work reached a natural boundary or because the session reached a limit. The first leaves an artefact somebody can check and a clean place to resume. The second leaves whatever was in progress, at a point chosen by capacity, with the state of the work known only to the session that is ending. The whole objective is to make the first happen before the second does.
The clearest seam is a phase change. Investigation to implementation, drafting to review, analysis to writing. The context that served the first phase is largely irrelevant to the second and is actively competing with it, so a fresh session seeded with the conclusions is sharper than a continuation even where the continuation would have worked.
The second is any point where a checkable output exists. A written finding, a completed stage, a change ready to be reviewed, a document that can be read. These are the places where a boundary costs nothing, because the artefact carries the state forward and the resumption reads it rather than reconstructing it. A split anywhere else requires a handover that describes work in progress, which is a much weaker object.
The third is the degradation signal, and it should be treated as an instruction rather than a hint. When answers stop referring to your specific material and start describing the general case, the session has stopped being useful and everything after that point is confidently generic. This can arrive well before any limit, and continuing because there is capacity remaining is the specific mistake.
What should not drive the decision is the token count or a percentage of the window. Those measure the state of the session and not the state of the work, and splitting on them produces a boundary in the middle of something. Where capacity is genuinely running out and no seam is near, the correct move is to create a seam — stop, write down the current state and findings deliberately, and start again — rather than to let the limit choose the moment.
One consequence for how work is planned. If splits should happen at checkable outputs, then a plan whose stages each produce one is a plan that can be split anywhere, and a plan with a single four-hour stage is one that cannot be split at all. The decomposition determines the available boundaries, which is another reason to do it before starting rather than during.
Split where the work has a seam, not where the session runs out, because only one of those places leaves something you can inspect.
Siddharth Sharma, Context Theory
Related questions
What if the work genuinely has no intermediate output?
Then create one, even if it is only a written account of what has been established. A task with no inspectable intermediate state is also a task nobody can review until it is finished, which is a problem independent of session length. The absence of a natural seam is usually a signal about the decomposition rather than about the work.
Does splitting lose momentum?
It loses the unwritten impressions the session had accumulated, which is a real cost and is why splits belong at seams rather than mid-thought. It does not lose anything that was written down, which is the argument for writing continuously rather than for avoiding splits.
METHOD
Every figure below carries its source and the date it was verified. Nothing on this page is asserted.
The numbers on this page.
| What | Value | Specific to |
|---|---|---|
| Sub-15-minute compliance — automated routing vs manual only | 62.5% vs 39.1% | Category-wide |
| Close rate — response under 5 minutes vs over 24 hours | 32% vs 12% | Category-wide |
2026 speed-to-lead benchmark · verified
Optifai speed-to-lead benchmark · n=939 companies · Q2 2025–Q1 2026 · verified
What is specific to this page.
| Kind | Claim | Check it against |
|---|---|---|
| Workflow | A session ending at a work boundary leaves an inspectable artefact and a clean resumption point, while one ending at a capacity limit leaves work in progress at a moment chosen by the tool with its state known only to the ending session. | Comparing what remains after a session ended at a stage boundary and one that ran out of capacity. |
| Response | A phase change makes a fresh session sharper than a continuation even where the continuation would work, because the earlier phase's accumulated context is largely irrelevant to the later phase and competes with it. | Performing a second-phase task in continuation and in a fresh session seeded with the first phase's conclusions. |
| Software | Token count and window percentage measure the state of the session rather than the state of the work, so splitting on them places the boundary in the middle of something, and creating a seam deliberately is preferable to letting a limit choose. | Checking whether the last session boundary corresponded to a completed unit of work or to a capacity threshold. |
| Buying behaviour | The decomposition determines the available split points, so a plan whose stages each produce a checkable output can be divided anywhere while a plan with one long stage cannot be divided at all. | Listing the points in a current plan at which an inspectable artefact would exist. |
Each row would be wrong on another industry's page. Where a sourced figure exists it is in the table above instead; these are the constraints that shape the work and do not happen to be numbers.
Start with the measurement.
Reading about a benchmark is not the same as knowing your own number. The audit produces yours, measured rather than estimated.
$497 · delivered in 5 business days · credited against month one