Stop Prompting, Start Directing: How to Put AI Agents to Work Without Losing Control

ChatGPT has 1.2 billion weekly active users, yet by OpenAI co-founder Greg's own account, people only become power users once they find three different things it can do for them. The gap between a tool that can do almost anything and people who use it for almost nothing is where the next wave of productivity, and of new businesses, will be won.

Why most people stall at the blank page

ChatGPT reaches 1.2 billion weekly active users, by the figure Greg, an OpenAI co-founder, gives in the source. About 300 million of them ask it health questions every single week. Measured by reach, the technology has arrived.

Greg puts a number on the gap from inside the company: people who discover three different things the assistant can do for them become power users.

The host names the newcomer's version of this problem directly: when people start using agents and automating for the first time, what do they start with?

A survey of more than 2,200 US businesses commissioned by Intuit QuickBooks in April 2025 found that 68% use AI regularly, but only 28% use it daily. Among small businesses on the Gusto platform that use generative AI, 22% use it regularly, 51% occasionally and 27% are still experimenting. More than 80% of those adopters use it for writing or research, and far fewer use it for higher-stakes work such as analysing financial or customer data. A U.S. Bank survey reported by Axios found that 68% of small business owners who use generative AI spend less than $50 a month on it.

Greg's design answer is that what people want from AI is simplicity, not interfaces full of buttons, sliders and model pickers. The host's practical question in the source is the right one to keep asking every few months: which use cases were unimaginable a year ago and are possible now?

Why agents work now: from chatting with a model to directing one

Figure 1. How the dominant AI product form has shifted as model capability expanded, from readable chat responses to autonomous coding agents to proactive cloud-based assistants.

Greg dates the eras of AI products by what the models could do. In 2022, a model was personalized enough that its responses were worth reading, which made a chat product make sense. At the end of 2025, he says, the coding agent era really started. In that shift, AI went from being 20% of software production to 80%.

Anthropic's engineering guide draws the useful line: workflows orchestrate a model and tools through predefined code paths, while agents let the model dynamically direct its own process and tool use.

A survey of agentic programming describes the loop: the model receives a prompt, gathers context from the workspace, breaks the task into subgoals, edits files or runs terminal commands, and feeds the tool output back into its reasoning until the task is done. Compiler errors trigger targeted edits, failing tests trigger debugging, and linter warnings guide style.

The SWE-agent research introduced an agent made of a language model and an agent-computer interface, a designed set of actions that lets it work on real software engineering problems.

This is why Greg can say he no longer writes software: he directs an AI to write it, giving it feedback and guidance. He adds that the models are now better at writing software than basically anyone he knows. Directing is a different job from typing code: you specify the outcome, set the checks that define done, review what comes back and decide what matters next.

A study of five repair agents across 500 real-world tasks found that they excel at simple fixes but struggle with logic-intensive bugs, often producing verbose or overfitted patches that merely satisfy existing tests. The same work found that collaborative human-AI workflows produced better organised, higher-quality test suites.

One pattern described on Martin Fowler's site is for the human to write the test scenarios and let the AI write the implementation that makes them pass; another adds a human review of each failing test before the AI writes code. Its exploratory evaluation found no clearly discernible difference when an agent was simply told to follow test-driven development inside its own loop.

Let's call this the director's loop: state the outcome, define the check, let the agent run, review the result, tighten the brief. Each turn of the director's loop moves more of the typing to the machine and keeps the judgment with you.

Greg is blunt about the limits of the familiar coding agent: it is task-based, micromanaged, lives on a laptop and stops when the laptop closes. In his words, that was never the AI people pictured five years ago. What he describes instead is proactive: an assistant that noticed an email he had ignored for days, researched the answer, drafted a reply and asked for his approval, which he says happened just before the recording. A proactive assistant, in plain terms, spots the pending item itself and brings you a proposal, rather than waiting for you to remember to ask.

Greg expects the way people use AI to be surprising relative to what we expect today.

The method: start small, then automate after five repetitions

Greg's first instruction to newcomers is almost provocatively simple: start small, type something random, press enter and see what happens. The point is to break the blank page with low stakes, not to get a brilliant answer.

A connector is the link that lets the assistant read from or act in another service, such as your inbox or calendar.

The rule that follows is the heart of the method: do a task by hand until you have done it five times, then automate it. Greg presents it as a software engineering maxim whose point is to keep your finger on the pulse of the problem, and the host immediately labels it the five times metric.

Programmers know an older cousin of this rule. Martin Fowler's book Refactoring quotes a guideline from Don Roberts: the first time you do something, you just do it; the second time you wince at the duplication but do it anyway; the third time you refactor. The book dates from 1999, and the idea traces back to a 1996 paper by Roberts and Ralph Johnson on evolving frameworks. One practitioner stresses that it is a guideline, not a law: sometimes the right moment comes only at four or five repetitions.

One reason given for the rule of three applies directly: you probably don't know as much as you think about the problem, and acting on a premature understanding causes a tremendous amount of rework.

Greg is firm that accountability stays with the owner: the customer signs up or not, pays or not. So owners should stay deep on the details of whatever they automate.

Anthropic's guidance adds the other half of the discipline: find the simplest solution possible and add complexity only when needed, which may mean not building an agent at all. It notes that agentic systems trade latency and cost for better task performance, that workflows bring predictability to well-defined tasks, and that optimizing a single model call with retrieval and examples is often enough.

A coaching firm case published by an automation vendor shows graduated trust in practice. Its lead follow-up bot drafted same-day replies, and the first touch was human-approved for the first two weeks before switching to auto-send with a clear opt-out. The vendor reports same-day replies rising from 48% to 96% within two weeks.

Figure 3. How a first agent is wired: the agent node connects to external services through permission-gated connectors, each set to the least permissive level that still gets the job done.

Setting up a first agent this weekend

Asked where someone with one hour should start this weekend to step up their agent game, Greg's answer is to sign up for ChatGPT dots, which he calls a really cool new form factor. He says it is powered by GPT6 Astra. It has its own cloud computer and can optionally hook into your local computer.

That cloud computer answers the laptop problem of the classic coding agent.

OpenAI's help center, as found for this article, states that ChatGPT agent is no longer available and points users to ChatGPT Work for longer, multi-step tasks. It describes a cloud browser that gives ChatGPT Work its own browser on a separate computer in the cloud, keeps running after you close your computer, and pauses when it needs input, sign-in or confirmation. Cloud browser is listed for paid plans in supported regions, excluding Free and Go.

Connections are what turn a chatbot into a worker. Greg notes that voice mode is now hooked up to tools: with email and calendar connected, you can ask what your schedule is today and get suggestions on how to free up time for deep thought. One of his friends used AI to produce CAD plans for a guest house, brought them to a contractor, and the house is now being built, a job that would normally require hiring someone.

OpenAI's documentation lists options running from Always ask, through allowing read actions and low-risk actions, up to Allow all actions, which it says carries elevated risk. A connection respects existing permissions: ChatGPT can reach only the content available to the account you connect, and a personal account does not open files that only your work account can see.

Recurring work runs through scheduled tasks, which ChatGPT can run once or on a repeating schedule. Active task limits depend on the plan: 3 for Free and Go, 5 for Plus, 10 for Business and Edu, and 15 for Pro and Enterprise. Free users can schedule a task no more than once per day, in flexible windows such as morning, afternoon or night. Older guides, such as a university help page, describe the connection path under Settings > Connectors, so menu names vary by account and date.

🛠 How to — Run a first agent on a recurring task, with you approving every action

1. Open the settings and select the app you need, such as Gmail or Google Calendar, then install or connect it if prompted. Settings > Plugins
2. Sign in with the account that already has access to the information, and read the requested services and permissions before approving.
3. Set the app's permission so that ChatGPT asks before reading information or making changes. Always ask
4. Ask ChatGPT to run the task on a schedule, and check that your plan's active task limit leaves room for it.
5. Review the first few runs, then adjust the prompt, tools or cadence if the results are too broad or miss context.

Tokens, compute and judgment: measuring what AI does for you

A token is the unit of text a model reads and writes, roughly a fragment of a word, and it is how model work is metered and paid for.

Greg describes an era of token maxing, when people blindly said to throw more tokens at problems, and calls that clearly not a primary objective. Yet he does not dismiss the number. If you are not trying to optimize tokens for their own sake, he says, you will probably see improved productivity correlated with increased token use.

Managers know the general principle: a measure that becomes a target ceases to be a good measure.

Greg's example of compute pointed well is scientific. OpenAI put a ton of compute into Navier Stokes, which he calls a big money problem: about 10,000 agents running for some number of days. He presents it as AI creating new knowledge. The Navier-Stokes equations describe how fluids flow, and open mathematical questions about them are among the most famous unsolved problems in mathematics.

The outsized results, in Greg's view, go to people who point compute at the right problem, with judgment and real thought about what really matters. Those same people, he adds, are willing to push their token budget high once they have found that problem.

Enterprise guidance on agents recommends defining key performance indicators such as accuracy, latency, cost and safety thresholds up front, then measuring behaviour continuously.

Asked whether his own use of AI is ambitious enough, Greg says the answer will always be no, because there is always more we can do. Having achieved one thing, in his framing, is a reason to dream a little bigger.

Small businesses and the coming renaissance of entrepreneurship

Greg's boldest prediction is economic: a massive renaissance of entrepreneurship, which he says is coming. His reasoning is that a founder now carries a PhD-level intelligence and a world-expert business consultant in their pocket for free, with more depth and compute for those who pay.

A small anecdote shows the scale of the change. At Stripe, Greg recalls, you emailed the data team a question and got an answer about a week later.

Adoption among small businesses is rising fast. It also found that 82% of small businesses using AI increased their workforce over the past year.

The barriers are concrete: in Gusto's data, the biggest are a lack of technical expertise and fears about data privacy and security, with around 40% of owners worried about privacy and security. The same survey found that businesses providing AI training see productivity gains upwards of 20%.

A garage door business featured by Workiz runs 50 automations covering estimate follow-ups, pre-job reminders, invoice follow-ups, review requests and receipts, which the company says replace the need for a team of employees. Another Workiz case, Gold Eagle Services, reports $23,758 in revenue in 30 days from conversations its AI answering handled, with replies to leads in under two minutes.

Read these cases the way an investor reads a pitch deck. One automation agency's cleaning-company case states outright that it is an illustrative composite whose figures are not verified client results.

Figure 4. Where durable advantage lives: as frontier models commoditize the base layer, defensibility moves up the stack into proprietary data, workflow integration, compliance, and evaluation infrastructure.

Greg also notes that recent Y Combinator batches are 60% B2B, meaning startups that sell to other businesses. Independent analyses land in the same range: one counted 65% B2B among 2025 companies, another found the share at 56.63% in 2026 as Industrials rose to 18.23%. An analysis of the Summer 2025 batch classified 141 startups, or 88%, as AI-native.

The host's own project shows the founder's version of this. Even with 12 years of documented files to draw on, each episode still takes at least 40 hours of work.

The host closes on a line worth adopting: everyone has homework to work on their ambition this weekend, not just on agents.

Building products that get better when the model does

Every builder working on top of a frontier model faces the same question: how do you build something that the next model release will not make obsolete?

Venture investors who study this put it bluntly: many startups find something frontier models cannot do today and build a company around that capability gap, and that gap keeps shrinking. Their advice is to build products and platforms that get better as the underlying models improve, which is what separates a durable company from a temporary bridge.

Better models do not destroy defensibility; they commoditize one layer of the stack and push the advantage into the other layers. Those layers are unique proprietary data, ownership of a multi-step workflow, a redesign of how the work gets done, deep integrations that frontier labs will not build, and the trust and compliance that large organizations require.

Data is the most cited moat and the most overclaimed one. Proprietary data only becomes a moat when it changes product behavior, through a closed loop in which use generates data, the data improves output, and better output attracts more use. The valuable records are the ones the public web does not hold, such as the corrections a domain expert makes to a machine draft or the outcomes recorded after a recommendation was followed or rejected. Check the contracts too: documents your customers upload usually remain their data under their contract, and rights that allow aggregation and model improvement across customers are rare.

If you must pick one moat to build first, one practical prioritization is workflow integration, because it creates switching costs immediately whether or not your model improves. Integration also produces the interaction data that a data flywheel later needs, so the order matters: no usage, no flywheel. A product that writes back into a customer's systems of record and is configured around their approvals is expensive to remove even when a better model appears, while a product that sits in a separate browser tab is not.

The least glamorous asset is often the most valuable: an eval set, meaning a library of real cases with known good answers that you run against every model or prompt change. It lets a team switch model providers in days rather than months, prove quality to customers and catch regressions before they ship.

The position defended in the source is that humans must remain in control and in charge, with the goals coming from people. It extends that to infrastructure: the systems through which AIs represent individuals and companies must be trustworthy and observable.

Watching an agent is harder than watching ordinary software, because AI systems are probabilistic and their behavior shifts with inputs, retrieved context, tool outputs and guardrail decisions. For the same reason, uptime and error rates say little about whether an AI system is doing good work. Enterprise guidance therefore pairs automated scoring, including using a model as a judge, with human-in-the-loop review, and keeps human oversight for exceptional situations.

ChatGPT offers levels that range from Always ask, through Allow read actions and Allow low-risk actions, up to Allow all actions, which its documentation labels as carrying elevated risk.

The source expects a world where people solve problems in fields such as material science that once required a very small set of specialized experts. It leaves open which breakthroughs become possible once that kind of AI is in everyone's hands.

Mistakes that sink agent projects, and what the evidence has not settled

A token is the unit of text a model reads and writes, so token count measures how much work the model did, not whether that work mattered. The source describes an era of token maxing, when people blindly threw more tokens at problems, and calls it clearly not a primary objective.

In a randomized controlled trial run by METR, 16 experienced open-source developers worked on 246 real issues from their own repositories, and when AI tools were allowed they took 19% longer. Before starting, the developers forecast that AI would cut their completion time by 24%. After the study, they still estimated that AI had reduced their time by 20%. The time went into prompting and reviewing AI output, which ate more than the coding it saved.

It measured early-2025 tools, and METR now states that these results are out of date, that it published new data on late-2025 tools in February 2026, and that the historical results no longer reflect current impact. That fits the source's dating of the coding agent era to the end of 2025, and its view that models now write software better than basically anyone the speaker knows.

A systematic analysis of five program repair agents across 500 real tasks found that they excel at simple fixes but struggle with logic-intensive bugs, often producing verbose or overfitted patches that merely satisfy existing tests. A controlled study of vibe coding found that agentic workflows deliver functionally correct code fast but can add implementation decisions nobody asked for, leaving untested decision points.

The recommended starting point is the simplest solution possible, adding complexity only when needed, which may mean building no agent at all.

The source describes the familiar coding agent as task-based, micromanaged and tied to a laptop that stops it when closed, and says that was never the AI people pictured five years ago. Cloud-based tools already break part of that pattern: ChatGPT's cloud browser runs on a separate computer and can keep working after you close your laptop, pausing when it needs sign-in or confirmation.

OpenAI's own guidance is to test a task's prompt in a regular chat first, then review the first few scheduled runs and adjust the prompt, tools or cadence if results are too broad or lack context.

The source itself leaves open what people will do once AI able to solve unsolved problems is in everyone's hands, and which breakthroughs will follow. Researchers note that accurately reproducing development workflows with agents is still an unsolved problem, even with the newest agentic tools.

Key points

• ChatGPT has 1.2 billion weekly active users, yet people become power users only after finding three distinct jobs for it.
• Do a task by hand five times before automating it, so you understand what you hand over.
• Token count is a proxy for delegated work, never the goal; point compute at the right problem first.
• Build products that improve as models improve, not ones that fill today's capability gap.

Action plan

1. Today, paste one email you have been dreading into ChatGPT and ask for a draft reply, without connecting anything.
2. Turn the logged task into a scheduled task, testing the prompt in a regular chat first and reviewing the first few runs.

Test yourself

Q. What separates a workflow from an agent in Anthropic's guidance, and which should you try first?
A. A workflow runs a model and tools through predefined code paths, while an agent directs its own process and tool use; start with the simplest solution and add complexity only when needed.

Q. What did METR's trial find about experienced developers using early-2025 AI tools, and how does METR treat that result now?
A. They took 19% longer with AI on their own repositories; METR now calls those results out of date and has published data on late-2025 tools.

Open questions

• How much of the at least 40 hours that each episode of the host's podcast still takes could AI actually take over, and what would owners do with the time recovered?
• How will businesses differentiate once building apps becomes commoditized, beyond understanding the problem and the users?
• What will people do once AI that can tackle unsolved math problems is in everyone's hands, and which breakthroughs will follow?
• How can a general-purpose assistant bring its users to a third use case, a step Greg himself calls a hard challenge?

Go further

1. Building effective AI agents \ Anthropic — Anthropic's guide to the difference between workflows and agents, and when not to build an agent at all.
2. Connecting and managing app accounts in ChatGPT | OpenAI Help Center — OpenAI's official steps for connecting an app account to ChatGPT and reviewing its permissions.
3. Managing app permissions in ChatGPT | OpenAI Help Center — The permission levels for connected apps, from Always ask to Allow all actions, and their risks.
4. Scheduled tasks | ChatGPT Learn — How scheduled tasks run, plus the advice to test a prompt in chat and review the first runs.
5. Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity - METR — METR's randomized trial on AI and experienced developer productivity, with its note on newer results.
6. The VC's Lens: How AI Is Rewriting the Rules of Defensibility — A venture investor's view of where defensibility moves as models improve, and why not to build for the gap.
7. Bob Nadler, Jr | The Rule of Three — The rule of three from refactoring, its origin and the reasoning for waiting before abstracting.
8. Observability for Generative AI and agentic AI systems | Microsoft Learn — Microsoft's case for why agents need AI-specific observability beyond uptime and error rates.