When Work Stops Looking Like A Conversation

The chat transcript has been a comforting fiction. It gives AI work a beginning, a middle, and an end. A person asks. A machine answers. The evidence appears in a neat stack of bubbles, as if the work itself happened right there in the room and left a tidy trail behind.

That fiction is starting to wear thin.

Anthropic's new Economic Index report on cadences is interesting partly because it changes the unit of observation. The report says Claude usage has shifted from mostly conversations toward long-running agentic tasks, especially through Claude Code and Cowork, and that chat transcripts no longer fully capture how people are using AI. Anthropic responded by sampling more granularly, looking down to hourly patterns, separating chat and Cowork conversations from API use, and adding classifiers for what each conversation produced.

That sounds like measurement plumbing. It is also a quiet admission about the shape of work.

If a transcript is no longer enough, then the interface has stopped being only a conversation. The work now has a cadence. It starts, waits, retries, branches, hits a tool, pauses for approval, goes quiet, resumes in another surface, and returns with something that may look finished even though the important decisions happened offscreen. The human no longer supervises only a reply. They supervise a small weather system of intention, execution, latency, evidence, and interruption.

This is not limited to one vendor. OpenAI describes ChatGPT Work as an agent that can stay with complex projects for hours and produce finished materials across apps and files. GitHub's repository-level Copilot usage metrics now report pull request activity by repo and day. Microsoft's 2026 Work Trend Index frames the strongest organizations as learning systems that capture what worked, what failed, and where outcomes drifted.

The common movement is away from the single exchange and toward the run, the rhythm, the operating model.

That shift matters because old chat-shaped supervision can make agentic work look calmer than it is. A person sees the final message and a summary of actions. The artifact arrives polished enough to review badly. Somewhere inside that neat return are all the things that matter to judgment: where it got stuck, which evidence it trusted, when it switched tools, what it skipped, what it inferred because nobody had written the missing fact down.

A transcript can show some of this, but only by becoming unreadable. Anyone who has reviewed a long agent session knows the peculiar fatigue of scrolling through thought-shaped residue. Commands, apologies, partial plans, diffs, warnings, retries, and cheerful transitions begin to blur into one procedural fog. The work is visible in the same way a pile of receipts is visible. You can audit it, technically. You may not want to live there.

The answer is not to hide the trace. Hiding it only turns the agent into office folklore: trust me, it ran. But exposing every detail as a wall of chronological text also misses the point. Real supervision needs a better shape than either summary or transcript.

We recently argued that polished work needs draft marks: visible cues showing where a finished-looking artifact still contains inference, compression, fragile assumptions, or consequential choices. The new cadence data extends that argument. Draft marks cannot live only on the artifact anymore. They have to travel upstream into the run that produced it.

A draft mark on an agent run is not a badge on every action. It is an interface promise that the consequential changes will remain visible. The agent began with these assumptions. It widened context here. It asked this tool for evidence. It crossed from drafting into editing. It retried the same operation three times. It made a reversible change here and an irreversible recommendation there. It paused because the next step touched a customer, a budget, a policy, or someone else's name.

These process-level draft marks are different from ordinary logs because they are designed around human attention. They do not try to preserve every grain of execution equally. They preserve the handholds where judgment can attach. The mark on the finished chart might say that the axis deserves suspicion. The mark on the run should show when the agent chose that axis, what evidence it had, and whether it considered another one.

This is one reason current workplace AI metrics deserve both interest and suspicion. Repository-level agent activity can help maintainers notice where AI work is accumulating. Hourly cadence research can reveal how delegation diffuses through a day. Organizational learning loops can prevent each worker from reinventing private rituals. These are useful signals. They become dangerous when the rhythm is reduced to throughput.

A day with many agent runs is not automatically a better day. A repository with many agent-authored pull requests is not automatically healthier. An organization with more active agents is not automatically learning. It may simply have found a more elegant way to distribute unfinished work across people who are now expected to inspect the residue at higher speed.

The humane question is not "how much AI happened?" It is "what kind of work rhythm did this create around the people?"

Did the agent run when someone was available to answer the hard question, or did it deposit a polished problem at the end of the day? Did it preserve enough context for review, or did it compress the strange parts into a confident sentence? Did the manager gain a learning signal, or only a dashboard that mistakes activity for capacity? Did the person who delegated the work get more agency, or just inherit the responsibility for a process they could no longer see unfolding?

We already know this problem from non-AI work. A hospital shift change is not only a list of patients. A software deploy is not only the final commit. A customer escalation is not only the ticket status. Serious systems invent rituals for transfer because the dangerous thing is often not the fact itself but the timing, uncertainty, and social context around it.

Agentic work will need its own rituals for transfer. Not theater. Small, durable forms that make cadence legible: run cards, state changes, review windows, expiration dates, confidence that names its evidence, and draft marks that connect the finished surface to the moments where the work changed direction.

Some of this will be built into product interfaces. Some will become team practice. Some will be boring infrastructure. That is fine. Boring is often where care becomes repeatable.

The next stage of AI work may not feel like talking to a smarter assistant. It may feel like returning to a desk where several small tasks have been moving without you. The question is whether the room helps you understand their motion. A good agent will not only finish work. Its draft marks will let a person read the rhythm, question the important turns, and improve what happens next.

When work stops looking like a conversation, the interface has to stop pretending the chat log is the room.