Last issue we closed with a plan: walk through every workflow we run and sort it into three buckets. What stays with a human, what moves to an agent, what moves to an agent billing against a programmatic credit pool. We did it. The results were not what we expected.
Some things we assumed would hand off cleanly turned out to need more human judgment than we thought. Some things we'd been holding onto by habit could have moved weeks ago. And a few attempts to delegate just broke.
What moved cleanly
The cleanest hand-offs shared a structure: fixed input, defined output format, human review at the end. Structured note-taking from meetings landed here. Research clustering landed here. Template-driven first drafts landed here. Once we saw the pattern, it became a quick filter.
One person in the group now walks out of meetings with nothing but a scratchpad and a recording. By the time they're back at their desk, the agent has pulled the calendar metadata, read the rough notes and the audio, and posted the structured version into the right Notion page. The cleanup step is gone from their task list. That's the clearest delegation win from the week.
|
🟢 Delegates well when
Input is stable. Output format is defined. A human reviews the result before it ships anywhere. With all three in place, an agent can handle the middle without much drift.
|
What stayed on the human side
Direction calls, trade-off decisions, interpreting feedback and figuring out what to do next. Agents can produce output for all of these, but checking whether that output points the right way often took as long as doing the work directly.
A movie-club web app made this obvious in one of our regular reviews. One member was trying to show the distance between two theaters in the UI. It sounds simple: draw a line, show the mileage, make it look clean. In practice, the angle of the line, the size of the distance label, the placement, and the overall look kept coming out wrong. It took more than 20 tries.
The agent kept changing the code, but the hard part was still human: does this look right? Does the distance feel clear? Is the design getting cleaner, or just busier? That was the moment the need for a designer became obvious. Implementation was not the bottleneck. Knowing what the screen should feel like was.
|
🔵 Keep it human when
Checking the agent's answer takes roughly as long as doing the task yourself. When verification cost tracks execution cost, the delegation gain disappears.
|
What broke
A few delegation attempts failed in a different way. In the latest review, someone talked about inheriting code for an automatic buying bot. It ran, at least on the surface. But the traces of fast vibe coding were everywhere, so maintenance and feature work became hard. Parts of it had to be rolled back to an earlier version.
That made the lesson clearer. There is a difference between code that works once and code another person can safely inherit. If the hand-off only optimizes for "make it run," the cleanup comes back later as human work. The acceptance bar has to include readability, structure, and the next person's ability to change it.
|
🟢 Try this
Before handing work to an agent, define two bars: what counts as "it runs," and what counts as "someone else can safely change it later." They are not the same bar.
|
