The Middle Manager Inside “In Progress”
The Middle Manager Inside “In Progress”
Say you've asked an AI agent to compare three competitors' pricing for a client proposal. You check in an hour later and the task says “In Progress.”
That tells you the work is moving. It doesn't tell you whether anyone has checked the prices against the competitors' own pages, or what happens if the comparison can't meet the standard you had in mind. The draft might have been sent back twice already, or never checked at all, and the status would read exactly the same.
On a human team, a good middle manager covers that stretch. They read the draft against what you asked for and send it back with a note when a price has no source. If a claim looks shaky, they bring in someone who knows the market better. And when a draft isn't getting any better, they stop sending it back and come to you. You rarely see any of that. You see a status, and then a draft worth reading.
That's the middle manager inside “In Progress”: the checking that happens while a task is still moving. When an AI agent is doing the work, that checking still has to happen. The question is who, or what, is doing it.
Two options that both fall short
Most conversations about AI work offer two choices. Trust the first answer the model gives you, or put a person behind every task.
The first is fast until the comparison goes out with one competitor's monthly price listed next to another's annual price. The second is safe, but it brings back the bottleneck you were trying to remove. If you have to recheck every line of the comparison yourself, you haven't saved much time.
There's a third option, and it's the one a good team already uses: build the checking into the work while it's moving. For the pricing comparison, that starts with writing down what a good one looks like, so a draft that misses can go back with a note and a claim that's hard to verify can go to a second reviewer. Someone also has to decide ahead of time what happens if any of that can't run.
Start with a standard someone wrote down
A check needs something to check against. For the pricing comparison, that might be: every price links to the competitor's own page and the date it was read, plans are compared at the same billing period, and any feature listed as included is confirmed on the plan page rather than taken from a review.
Without something like that, a check is someone's impression of the draft. With it, a reviewer, human or AI, can say exactly what missed and why.
The standard doesn't have to be the same for every piece of work. You'd read a new hire's first client deliverable more closely than one from someone who's done the job for years. A proposal going to a client deserves a higher bar than notes for an internal meeting. Whatever is doing the work, you should be able to see the bar it's holding the work to, and change it when it doesn't fit.
Sending work back needs an edge
Once there's a standard, a miss needs a response. The obvious one is to send the work back with a note: two prices have no source, and the third competitor's starter plan is only sold annually.
There's research behind the idea that a second pass helps. In Self-Refine, a 2023 paper by Aman Madaan and colleagues, a model drafted an answer, critiqued it and revised it, with no extra training. Across seven tasks, people and automatic metrics preferred the revised outputs to single-pass answers, and task performance improved by about 20 points on average.
A loop also needs an edge. If the comparison keeps coming back with the same unsourced price, sending it back again isn't working, and something has to stop the loop. Usually that's a set number of attempts. After the last one, the work might come to a person, be held, or go through as it is. Only the first two keep that price out of the proposal.
A second reviewer for the harder calls
Sending work back only helps if whoever reviews it can see what's wrong. With AI, that's a real question.
Large Language Models Cannot Self-Correct Reasoning Yet, a paper by Jie Huang and colleagues presented at ICLR 2024, tested what happens when a model tries to correct its own reasoning with no outside feedback. The authors found that models struggled to fix their answers that way, and sometimes did worse after trying.
The Self-Refine paper points the same way once you look past its average. Its biggest gains came on tasks like rewriting a reply in a conversation to be more engaging. On math problems, the revised answers barely changed, and the authors traced that to the model not spotting its own mistakes. In 94% of the math cases, ChatGPT's feedback on its own work was that everything looked good. When something outside the model flagged a wrong answer, the gains were much bigger.
That's why a manager brings in someone else when a claim is hard to verify. In the pricing comparison, the model might be confident that a competitor's starter plan includes unlimited seats because a review said so. A second reviewer who knows to open the plan page, or who has a source the first one didn't use, is more likely to catch that than the same model rereading its own work.
Not every piece of work needs that second look. Notes for an internal meeting can skip it. The pricing comparison, with a seat count taken from a review and two send-backs already behind it, shouldn't. So ask what triggers a second reviewer, and what happens when none is available.
When the check itself breaks
What happens when the checking step itself fails? Say the scoring errors out halfway through the comparison, or the second reviewer can't be reached.
There are two reasonable designs. The work can pass through, so a broken check doesn't hold up everything behind it. Or it can stop, so nothing goes out unchecked. Passing through keeps work moving, but it means some work reaches you without the check you were counting on. Stopping is the safer choice for a comparison going to a client, and a frustrating one for everything that didn't need it. A system might do both, letting routine work through and stopping when a review was required.
Five questions to ask before you trust “In Progress”
Before you treat a status as a sign that work is ready, ask these about any system doing work for you, including ours:
- Which work gets checked, and which skips the check entirely?
- What standard is the work held to, who set it, and can you raise it for work that matters more?
- How many times can work be sent back, and what happens when it runs out of attempts?
- When does a second reviewer look at the work, and what happens if none is available?
- What happens when the check itself fails? Does the work pass through, or stop?
If the answers are clear, you know what “In Progress” means for your comparison. If they aren't, the status is telling you less than it seems to.
The other half of a manager's job
A good manager does one more thing a checklist can't. Over time, they help the people on their team get better at the work. The analyst who kept putting monthly and annual prices side by side learns to catch it before anyone else has to.
We built that into Supanova. Atoms are paired with mentors, and the pairings change as their performance does. An atom is promoted as its work earns it. We wanted an autonomous workforce that develops, the way a good team does.
None of that replaces the five questions, or your own read of the finished comparison. When you green-light a project in Supanova, your team runs it end to end, and finished work comes back for your review. Nothing ships without you: you read the comparison, and you decide whether it goes to the client.
See what an autonomous workforce can take on for you.