A founder I worked with recently gave his brand-new AI agent a simple instruction before going to bed: "Find me 90 leads while I sleep." He woke up to eight, and four of those were the wrong people entirely.
His first assumption was that the agent wasn't smart enough. That is almost always the wrong place to look. The models are capable. What broke was everything around the model: the agent had no idea how to actually succeed at the job it was handed.
Over the past few months I have watched a lot of people hit this same wall, and the failures fall into three specific gaps. Close all three and an agent goes from an unreliable intern to something you can lean on. Here is each gap, and how to fix it.
1. It has no runbook, so it reinvents the job every time
Left on its own, an agent does a task by thinking it through fresh on every run. It leans on whatever it can remember and whatever the model can reason out in the moment. That works once. Then tomorrow it forgets the specific thing it learned yesterday, or the correction you gave it, and you are back to the start.
The agents that actually work do something different. The first time through a job, the agent, with your input, writes a runbook: the exact steps, the edge cases, the things to avoid. Every run after that begins by reading that runbook and following it. When it learns something new, it updates the document.
The difference in outcome is large. One agent wakes up at 2pm, reads "here are the 30 steps to post today's comments," and executes them. The other wakes up at 2pm and tries to recall how it did this last week. Give your agent a place to write down how the job is done, and tell it to read that before it starts anything.
2. It doesn't ask enough questions before it acts
Tell a person "grow my LinkedIn" and they ask you ten questions before touching anything. They want to know the audience, the tone, and whether the goal is posting, commenting, or making connections. Most agents skip all of that. They hear the instruction, decide they understand it, and charge off on their own assumptions. Those assumptions are usually wrong, and you find out only after the work is done badly.
The fix is to make the agent interrogate you up front. Before it runs the job, it should collect the things it cannot possibly guess: your voice, an example of a good outcome next to a bad one, the specifics that separate "this is great" from "this is not what I meant." The founder who woke up to eight bad leads was never asked a single question. Five minutes of gathering context first would have changed the whole overnight run.
One practical trick: ask in small batches with clickable options rather than one long wall of text. People will happily tap through twenty quick questions. Almost no one will sit down and type five paragraphs into a chat box.
3. It doesn't have the tools it needs
The third gap is the most concrete of the three. An agent asked to send email needs a way to send email. An agent asked to research prospects needs a data source. Very often the agent is willing and configured correctly, and it still cannot do the job because the connection was simply never made.
Good agents notice this and say so out loud: "I need to send outreach. Do you have an email tool? You do, but it is not connected yet, so let's link it." Wiring in the accounts and data sources is the unglamorous part of the work, and it is exactly what separates a slick demo from a workflow you use every week.
Put the three together
Something shifts when all three are in place. An agent with a runbook, real context from you, and the tools it needs gets roughly 90% of the way to a genuinely useful result. The last stretch is a short check-in, about once a day: anything blocked, anything to adjust, here is what I would do next.
That picture is very different from the one people expect. The story that spreads online is that you kick off a job at night and wake up to finished magic. The truth looks more like human agent collaboration: a heavy first stretch of teaching and calibrating, then a light daily rhythm where the agent does the work and you steer it. Less magical, far more dependable, and dependable is the version that actually earns a place in your day.
None of this asks for a smarter model. It asks you to treat the agent like a new hire rather than a vending machine. Write down how the job is done. Ask before you act. Connect the tools. Do those three things and your next "while I sleep" instruction has a real chance of being finished when you wake up.