Every honest account of running an agent eventually has to say what it does not hand over. Mine is about a fifth of the work, and the reasons cluster into four kinds.
1. Anything where being wrong is expensive and silent
The agent is good at tasks where a mistake shows up immediately: a broken build, a failed test, a page that does not load. It is bad at tasks where a mistake looks like success.
An example from my own week: renaming a column in a schema. The migration runs, the app starts, everything is green, and one report has been quietly wrong since Tuesday. Nothing in the run said anything was wrong, because from the agent's point of view nothing was.
I still do those tasks, I just do them with the agent watching me rather than the other way around.
2. Anything that needs taste
Writing the first draft of a post is not the same as deciding what the post is about. The agent can produce something competent and slightly off, and then I spend longer fixing the tone than I would have spent writing it. The line for me: if the output needs to sound like a specific person, that person writes the first version.
3. Anything touching another person's account or data
Not because it would fail, but because a failure is not mine to absorb. Support replies, emails under my name, anything that reaches a third party: the agent drafts, I send. The draft is still most of the work, and the send is fast.
4. Anything I cannot verify in under a minute
This one took me longest to accept. If checking the work takes longer than doing it, delegation is theatre. A ten minute check on a two minute task is a loss even when the agent gets it right.
What the boundary is actually made of
Notice that three of the four are about verification cost, not capability. The model could probably do most of the tasks I keep. I keep them because I cannot cheaply tell whether they were done correctly.
That reframing has been the most useful thing I have learned: the question is not "can the agent do this", it is "can I tell, quickly, whether it did". Everything I have automated passes that second test first.
I am curious where other people draw the line, especially for anything that reaches a customer. My guess is that support is where most people are more conservative than they expected to be.
Top comments (0)