The Manager Loop: How to Supervise AI Agents on Multi-Day Projects
The manager loop technique lets one agent interview you, then delegate work to execution agents. Here's how to structure it for complex tasks.
What is the manager loop?
The manager loop is a way of structuring AI agent work so that one agent handles the thinking and coordination while separate execution agents handle the doing. Instead of typing detailed instructions into a chatbot and iterating step by step, you describe a goal in plain language to a manager agent. That manager asks clarifying questions, breaks the goal into parallel workstreams, and hands each piece to a dedicated execution agent. You approve direction at a high level and step back in only when something needs a decision only a human can make.
The technique isn’t new. Similar setups have gone by names like “chief of staff agent” for a while, using one agent to manage others on your behalf. What’s changed is that newer models, including OpenAI’s Astra, are capable enough at long, multi-step computer use that the manager loop finally has something powerful to delegate to.
TL;DR
- The manager loop splits AI work into two roles: a manager agent that interviews you and plans, and execution agents that carry out the actual tasks.
- It exists to solve a real bottleneck: agents can now do work so complex that humans struggle to write a single prompt that fully specifies it.
- A household move is used as the illustrative example: over 20 hours of administrative tasks (housing search, school comparisons, doctor lookups, DMV appointments, utility transfers) with dependencies between them.
- Simple, contained tasks (find a doctor near an address, compare utility plans) don’t need a manager loop at all. You can hand those directly to any capable model.
- Complex, entangled tasks where one decision (like a moving date) cascades into a dozen others are where the manager loop earns its keep.
- The manager agent’s job is to ask the right questions up front (budget, priorities, which accounts it can access, where to stop and check with you) so it can turn vague goals into concrete, delegable jobs.
- This mirrors how people already work with skilled colleagues: you approve the big decisions and let them figure out execution, rather than dictating every click.
Other agents start typing. Remy starts asking.
Scoping, trade-offs, edge cases — the real work. Before a line of code.
Why do agents need a manager now?
For most of the history of AI chatbots, work was iterative. You’d ask for a piece of output, a paragraph, a checklist, a snippet of code, review it, and ask again. That loop kept humans in the driver’s seat for every micro-decision because the models weren’t reliable enough to string many steps together on their own.
That started to shift with coding tools in the back half of last year, when agents went from writing individual functions on request to being handed an entire coding job and working through it independently, hitting errors, trying alternate approaches, and continuing until the job was actually done. The same shift is now happening in general knowledge work with models like Astra, which can move across a browser, a calendar, a document, and a form-filling website in one continuous run instead of stopping after each answer to hand control back to you.
The catch is that as these agents get more capable, the limiting factor stops being “can the AI do this” and becomes “can I even describe this well enough.” A household move is a good stress test: it involves finding housing, comparing school districts, locating a pediatrician, prepping DMV paperwork, transferring utilities, and juggling dependencies where a shifted move-in date reshuffles half the other tasks. Writing a single prompt that captures all of that, plus every branch and contingency, is not realistic for most people. That’s the gap the manager loop is built to close.
How do you set up a manager loop for a complex project?
The structure has three parts.
1. State the outcome, not the steps. You tell the manager agent something like “I need to move my household to a new city by a specific date and I don’t want to lose a week of my life to forms.” You are not specifying which website to check first or how to format a spreadsheet.
2. Let the manager interview you. A good manager agent responds with the questions a competent human project manager would ask: Who’s moving? What’s the budget? What matters in a neighborhood? Are schools, doctors, vehicles, or pets involved? Which decisions are already locked in? Which accounts or documents can the agents actually use? And critically, where do you want it to stop and come back to you rather than proceed on its own? Answering these questions is far easier for most people than trying to anticipate them all in a single upfront prompt.
Remy doesn't write the code. It manages the agents who do.
Remy runs the project. The specialists do the work. You work with the PM, not the implementers.
3. Let the manager delegate to execution agents. Once it has your answers, the manager turns them into discrete jobs and assigns them to execution agents that can run in parallel where possible. Housing search can start immediately while school research runs alongside it. A doctor search can kick off once a shortlist of neighborhoods clears your review. DMV prep can begin once the destination, vehicle details, and move date are confirmed. The manager tracks all of these threads so you’re talking to one coordinator instead of juggling fifteen separate agent sessions yourself.
What tasks belong in a manager loop, and which don’t?
Not everything needs this structure. If a task is self-contained (find three pediatrians near a new address, verify their contact details from official sources, and summarize them for review) you can hand that directly to a capable model without any management layer. That kind of task is well-defined enough that a single clear prompt gets the job done, whether you’re using a model from OpenAI, Anthropic, or elsewhere.
The manager loop earns its place when tasks are entangled: when one piece of information changes several other workstreams, when there are approval gates you want control over, and when the sheer number of moving parts makes a single prompt impractical to write. A move is a good example because a shifted timeline cascades into housing, schools, medical care, and vehicle registration simultaneously. Complex multi-day or multi-week projects with many interdependent parts are the natural home for this technique, not quick single-purpose lookups.
Is agent supervision necessary, or can you just let it run?
Agent supervision, meaning some structured way of checking that an agent’s work stays on track and trustworthy, is part of what makes the manager loop viable rather than reckless. The framing here isn’t “hand over a credit card and walk away.” It’s closer to how you’d work with a capable colleague: you approve the important calls and let them handle the mechanics of getting there, but you still expect visibility and the ability to intervene.
Practically, that means designing the manager agent to flag decision points rather than barrel through everything autonomously. You decide in the interview stage which milestones need your sign-off (a final housing choice, a specific spending threshold) and which the manager can just execute and report back on. As agents take on longer and more complex runs, this kind of built-in checkpoint structure matters more, not less, because the cost of an unsupervised mistake scales with how much ground the agent covers in a single run.
Frequently Asked Questions
What’s the difference between a manager agent and an execution agent?
A manager agent handles planning, questions, and coordination. It doesn’t do the hands-on work itself. Execution agents are the ones that actually browse websites, fill out forms, pull data, and complete the concrete tasks the manager assigns to them.
Do I need multiple AI tools to build a manager loop?
No. The technique can be run within a single capable agent platform that supports spinning up sub-agents for delegated tasks, rather than requiring separate apps stitched together.
What kinds of projects benefit most from a manager loop?
Multi-part projects with interdependent steps and a timeline, such as a household move, a business launch checklist, or any project where changing one variable (a date, a budget, a location) affects several other tasks at once.
How do I know when to let the manager proceed versus step in myself?
Remy is new. The platform isn't.
Remy is the latest expression of years of platform work. Not a hastily wrapped LLM.
Decide this during the initial interview stage. Tell the manager agent explicitly which categories of decisions need your approval (spending limits, final choices between options) and which it can handle and simply report on afterward.
Is the manager loop only useful with the newest, most advanced models?
The concept works with less advanced models too, but it becomes far more valuable as models get better at sustained, multi-step computer use, since that’s what allows execution agents to actually complete delegated work without constant hand-holding.
