the rise and fall of the bifurcated screen

(insert spinner verb)

from roughly mid-2024, once models capable of sustained, high-quality iteration on documents and data were widely available, through late 2025, this arrangement was common enough to notice in many workplaces. it was especially visible where agents that could carry work forward without continuous prompting had not yet become ubiquitous. imagine a café on a weekday afternoon, its tables crowded with laptops, coffee cups, and charging cables whose reach determines the seating plan. suppose half the open screens share an arrangement: a document or a spreadsheet on one side, an ai conversation on the other. the work varies in nature, but its occupants have made a similar concession: they have accepted a smaller working area to keep somewhere to ask about it. call this the bifurcated screen: a workspace divided between a task and an ai conversation kept open to help develop, question, iterate on, and revise the work in progress.

at one table, a writer is preparing a proposal. a paragraph crosses the divide and returns considerably more certain of itself than its author had intended. the writer asks for the qualification back, specifies which sentence should remain untouched, and tries yet again. while the response takes shape, a small spinner turns beside a verb announcing that something is happening. ‘thinking’, ‘considering’, ‘pondering’; that waiting interval has acquired a budding vocabulary all of its own. the writer returns to the document, notices the response arriving, and off he continues looking across again. some suggestions survive, while others make the problem easier to identify precisely because they are unsuitable. explaining why the paragraph should remain tentative gives the next request a clarity the first lacked, so essentially, the conversation earns its place through that to-and-fro, including the exchanges in which very little generated wording reaches the page.

a friend opposite begins telling a story just as another answer appears. the writer asks for a moment, reads it, sends a clarification, and looks up again. the friend has returned to their own respective screen. both friends have small pauses in their work, but, regrettably, the pauses keep missing each other.

for the proposal’s argument, close conversation may be worth the divided attention. its supporting comparison table, however, puts forth a different kind of assignment. the writer knows which sources to use, and which columns belong in the table, including what to do with conflicting figures. those requirements can be supplied together, with discrepancies collected for review as opposed to settled through an interruption at every row. the writer keeps working on the argument while the table begins taking shape elsewhere. the first attempt might naturally need its rightful clarifications, and its instructions adjusted, but there is now enough direction for the assignment to continue between visits. the agent has access to the specified material, and any accompanying email or correspondence must remain a draft until the writer authorizes its sending.

somebody drags the divider a little farther right because the paragraph needs more room. after a while, the conversation is closed altogether, manifesting as the proposal now comfortably occupying full display real estate, while the comparison continues without it.

“awaiting approval”

background agents make that aforementioned forgoing of the opened conversation possible. OpenAI’s february 2026 Codex app announcement described scheduled automations whose results arrive in a review queue, allowing people to return after the assigned work has run. Cursor’s Automations, introduced the following month, can start agents on schedules or in response to events. in that arrangement, “always-on” describes an available process with defined triggers; a model need not be continuously thinking for work to begin without another typed request.

Luke Wroblewski described a similar retreat of chat in 2025, as agents took on intermediate coordination and people returned to examine the output. the visible conversation can recede while assistance continues, although the person’s experience depends on how often the work still asks them to return. a hidden panel can still occupy an afternoon quite effectively, if it incessantly asks for another instruction or leaves its owner unsure whether anything is actually progressing. the writer keeps checking, and the document has grown wider, but the friend opposite still has to hold their place in the story.

the comparison task in this café has been arranged differently. the writer has specified what can proceed and what must wait. sending the email requires an approval enforced by the system, such checkpoints are already supported in products including OpenAI’s workspace agents, where permission can be required before actions such as sending an email. a completed table and an approval request give the writer something definite to return to. it’s either the sources still needing examination or a discrepancy requiring further investigation, and perhaps even that email may need fewer exclamation marks before it’s sent. this merely allows the writer to choose a suitable moment to perform that review when the deadline permits, and is not a way to render work a nuisance or a formality.

granted, an ambiguous argument may still deserve half the screen tomorrow or the day after, while a recurring comparison may need only a brief and a later examination. experience supplies a more particular answer than either keeping the assistant permanently visible or insisting that it disappear. the bifurcated screen can lose its permanent place without losing its usefulness. its owner can reopen the conversation when the work calls for one. what changes is the expectation that every useful task must advance in step with the person’s next message. in this manner, a few scattered pauses can begin to resemble an interval.