Botsitting Is Eating the Hours Agents Were Supposed to Save
Workers spend hours a week botsitting AI agents: feeding context, catching mistakes, cleaning up. Autonomy levels and one inbox fix why.
Founder, Task Machine
Sol Rashidi runs strategy for a data security company and teaches at Harvard Kennedy School. She also ran four AI agents at once, until she fired two of them.
"I just fired half my agents because they were unreliable," she told Business Insider in July 2026. The agents were supposed to free up her time. Instead they demanded constant supervision: context that drifted and needed correcting, mistakes that needed catching, output that needed cleaning up before it was usable. "I don't have the time to babysit agents and keep course correcting the context," she said. Her fix was to hire human virtual assistants for some of the work instead.
Rashidi is one of a growing number of people a Glean report calls "botsitters" — workers who spend hours every week feeding AI context, debugging its mistakes, and cleaning up after it. The report puts the average at 6.4 hours a week, nearly a full working day, for white-collar workers using AI at their jobs.
Botsitting is not a training problem or a model-quality problem. It is what happens when an agent's output has no defined path to being trusted, so a human becomes that path by default, on every run.
Why babysitting is the default, not the exception
An agent that is told to "keep going until it's done" has no natural sense of when to stop and ask. A person doing the same recurring work eventually notices they are repeating themselves, or that a task has quietly become bigger than it should be, and pauses. An agent left alone does not have that instinct unless the workflow gives it one.
Without an instinct to stop, the only remaining safeguard is a human checking in. So people check in constantly: re-reading output before it goes out, re-supplying context an agent lost between runs, and watching a terminal to see whether the last hour of work is actually usable. None of that is a single failure. Each check is small and reasonable on its own. Added up across four agents and a full week, it is close to a full day of work that looks nothing like the work the agent was hired to do.
Rashidi's fix — fewer agents, more humans — is a rational response to that math. It is also the wrong lesson. The problem was never the number of agents. It was that nothing in the setup separated "the agent is doing something I don't need to see" from "the agent just did something that needs my judgment before it counts as done."
The fix is structural, not fewer agents
Botsitting disappears when supervision stops being a habit and becomes a property of the workflow itself. Two things have to be true for that to work.
The first is that every agent needs an explicit, named answer to "how much can this agent do before someone has to look." Task Machine calls this an autonomy level, and it is a ladder rather than a single on/off switch:
| Level | What the agent can do alone |
|---|---|
| Supervised | Only works the tasks it is directly assigned. No delegation, no new agents, no workflow starts. |
| Balanced | May attempt any consequential action, but each one waits for approval before it takes effect. |
| Autonomous | Delegation, task assignment, and workflow starts happen directly. Creating new agents or workflows still waits for approval. |
| Full | Every action applies directly, including creating new agents and workflows. |
An agent starts at the level you set, and you raise it a step at a time as it earns the room — never the other way around. That single setting replaces a habit of watching everything with a decision made once, per agent, that can be revisited when the agent proves it deserves more room.
The second is that the review the agent's work still needs has to land somewhere specific, instead of wherever you happen to be looking. Task Machine organizes work around three surfaces: chat to direct agents and fan work out into tasks and workflows, tasks for the detailed back-and-forth on one piece of work, and the inbox for everything that needs a decision — approvals, questions, proposed work, exceptions. Botsitting is what happens when there is no third surface, so the review work leaks into every terminal, every chat scrollback, and every "let me just double check this" moment during the day. An inbox turns that leak into a queue you clear on your own schedule.
What botsitting tasks actually need
| What botsitting looks like | Why it happens | What replaces it |
|---|---|---|
| Re-reading every output before sending it | No verifier or approval step decided in advance what "good enough" means | A verifier check or an approval node the workflow enforces before anything ships |
| Re-supplying context an agent lost between runs | The agent has no durable memory of the work it already did | Memory and task history attached to the work, not the chat session |
| Watching a terminal to see if a long job is still on track | No inbox item exists for the moment the job actually needs a decision | An inbox item that only appears when the workflow reaches a real decision point |
| Deciding case by case whether to let an agent keep going | No autonomy level is set, so every action is judged individually | A named autonomy level per agent, raised deliberately over time |
| Fixing the same class of mistake repeatedly | Nothing records what went wrong last time | Step-level workflow history you can read back and correct from |
Read down the list and the pattern is the same one Rashidi ran into: every row is a decision that had no home, so it defaulted to "whoever is watching, right now." Giving each one a home is what turns babysitting into review.
The honest limitation
None of this makes the review work disappear. An autonomy level does not mean an agent stops needing judgment applied to its work — it means the judgment happens at a decision point you chose, instead of continuously. And raising an agent from supervised to autonomous is not instant. It happens after you have watched it make a run of good calls at the level below, which takes real time before it pays off.
The honest version of this fix is not "agents that need zero attention." It is agents whose need for attention is legible, bounded, and routed to one place, instead of spread across every terminal and chat window you have open.
Where this fits in Task Machine
This is the operating model Task Machine is built around. Every agent carries an autonomy level, from supervised up to full, and you move it as the agent shows it can be trusted with more. Work runs across the three-surface workflow — chat to direct, inbox to approve, tasks to dig in — so the moments that need your judgment surface as a specific inbox item with the context attached, instead of a vague feeling that you should go check on something.
The difference this makes is not that agents need no supervision. It is that supervision stops being a background hum you can never fully turn off, and becomes a queue you clear with intent, on a schedule you control.
If you are running agents and spending more of your week botsitting them than benefiting from them, join the private beta on the waitlist.