Greg Isenberg's AI Agents: 6 Jobs They Do for His Team

What happened

Most founders spend their mornings in Slack. Greg Isenberg spends his sending voice notes to software. The CEO of Late Checkout — a studio that builds and backs small internet products — says his AI agents now take the first pass on almost everything the company makes, which is why he ends up exchanging more words with them than with the people on payroll.

The interesting part isn't the subscription. It's the assembly line: six named jobs, each handed to an agent, plus a few blunt rules about which model does what and when a human gets to override it.

One caveat before we go further, because the numbers matter. The original post describes a team of "50+ people in 12 countries." Public company listings put Late Checkout in the 11–50 employee range, so treat the exact headcount as fuzzy. The workflow is the part that's documented — Isenberg has written about it publicly on his AI agents page — and the workflow is what you can actually copy.

Six jobs that no longer start with a meeting

Here's the pipeline, in plain terms:

1. Idea stress-testing. He records a voice note. An agent fires questions back at the idea before a single line gets built. Think of it as the friend who won't let you buy the domain until you've explained who's paying for it.

2. Research. Agents read competitors' sites, prices and user reviews. The prompt is unusually specific: find the ten complaints that show up most and quote the users saying them. Not "summarize the market" — a tight, countable output you can argue with.

3. Design. Claude Code produces four versions of every screen before any code exists. Four, not one, because the first idea is rarely the good one.

4. Build. Agents work in parallel overnight. He checks what shipped over coffee. That's the entire morning standup.

5. Testing. Agents click through every flow three ways: as an impatient user, as a confused user, and as someone on a phone.

6. Bug fixes. An agent opens a fix in under 10 minutes. A human just approves it.

Three rules that keep the output usable

Anyone can throw prompts at a model. The rules are what stop the whole thing from turning into slop:

• Cheap model for sorting, the strong one for anything that ships. Tagging, sorting and categorizing is grunt work — a low-cost model handles it fine. Opus, Anthropic's top-tier Claude model, gets the work that goes out the door.

• A second model scores every output from 1 to 10. Anything under 8 goes back. It's an editorial pass, automated. Nothing mediocre quietly reaches the user.

• Every agent mistake becomes a new skill. You write the correction down once, and that specific screw-up doesn't happen a second time.

And the humans? Creative feedback, relationships, and telling the agents exactly what they got wrong. Which, honestly, is most of the job.

What it means for you

You probably don't run a product studio. That doesn't matter. The shape of the workflow — one narrow job, a numbered output, a second pass that scores it — works anywhere.

At home: the "ten complaints" prompt beats a to-do list

Say you're planning a trip or remodeling a kitchen. Instead of asking an assistant to "help you plan," ask it for the ten things people complain about most after doing this exact thing, with quotes. You'll get a checklist of real headaches instead of a generic guide. That's the research step, minus the company.

At work: one reviewable output instead of five open tabs

The magic isn't that an agent reads competitor pricing. It's that the deliverable is a numbered list with quotes attached. Your manager can approve or reject a list in 30 seconds. Nobody can approve "some research." Narrow outputs are what make AI work actually usable in a team setting.

For a small business: research you'd never pay for

A one-person shop can't afford a junior analyst to comb through reviews for two days. An agent doing it in minutes changes what's economically possible. You'll know which complaints are worth building against before you spend a dime, which is the difference between launching something people want and launching something you liked.

For studying: let the agent grill you

Isenberg's first step is an agent that attacks the idea before it gets built. Flip that on yourself: paste in your notes and ask the assistant to ask you five hard questions you can't answer yet. It's a study partner that doesn't get bored and doesn't flatter you. The gap between "I read it" and "I can defend it" closes fast.

For creative work: four versions, zero ego

Asking for four versions of a screen — or four headlines, four thumbnails, four openings — kills the instinct to fall in love with draft one. You pick, you combine, you move on. The cost of trying version four drops to nearly nothing.

For income: the overnight build

Agents working in parallel overnight means the bottleneck becomes your judgment in the morning, not your hours at the desk. Even a trimmed version — one agent drafts while you sleep, you review over coffee — changes your output per week. That's a side project that ships instead of a folder of half-finished ideas.

How to try it right now

You don't need six agents or a company. You need one job and one afternoon.

Step 1 — Start free. If you just want to see what these models are good at before spending anything, run your first prompts through the free tools at mykreatool.com. Try the research prompt there first; it costs nothing and tells you whether the output is worth building a workflow around.

Step 2 — Pick exactly one job. Research is the easiest starting point because the output is text you can check. Don't try to automate ideas, design and build on day one.

Step 3 — Write the prompt like a brief, with a number in it. Copy the structure: "Read these competitor pages and reviews. Find the 10 complaints that show up most. Quote the users." A number plus a quote requirement is what turns a vague summary into something you can act on.

Step 4 — Add the scoring pass. Run the output through a second model with one instruction: score this from 1 to 10 for usefulness and send anything under 8 back with reasons. This single step removes most of the junk.

Step 5 — Match the model to the stakes. Cheap, fast model for sorting and tagging. A strong model — Opus, in Isenberg's setup — for anything that ships. And Claude Code when you get to screens and code: four versions before anyone builds.

Step 6 — Keep a mistake log. Every time an agent gets something wrong, write the correction as a rule you paste in next time. Ten lines in a notes file is the whole "skill library."

Step 7 — Review in the morning. Let the batch run overnight, then spend your coffee checking what came back. Approve, reject, correct.

If you want the fuller picture of how this fits into building a company, Isenberg's AI startup notes and the Late Checkout site are both public, and this breakdown of his framework walks through the growth side.

Upsides and what changes

Your first draft stops being the expensive part. Writing, sketching, researching, testing — the first pass used to eat days. Now it eats minutes, and your time goes into judgment.

Quality control becomes a system, not a mood. A score threshold and a mistake log mean standards don't depend on how motivated you feel on a Tuesday.

Small teams punch above their weight. A handful of people with a tight agent pipeline can cover research, design reviews and QA that used to need specialists.

Your job description shifts. Isenberg's humans do creative feedback, relationships, and correcting the agents. That's a promotion, not a demotion — but it does mean the skill you most need is the ability to spot what's wrong with a good-looking output.

Speed changes your appetite. When four versions of a screen exist before lunch, you stop rationing experiments. That's a real strategic advantage, not just a time saving.

Limitations

Be skeptical of the tidy version of this story. Headcount is the obvious one: the post says "50+ employees in 12 countries," while public listings place the company in the 11–50 range, and "he talks to AI more than to his team" is a colorful line rather than a measured statistic. Nobody publishes the cost side either — multiple models in a loop, plus a scoring pass on every output, means you're paying per run, and iterating gets expensive quietly. Multi-agent orchestration also fails in boring ways: agents loop, they confidently agree with each other, they miss context a two-minute conversation would have fixed. And the approval step is still a human bottleneck — if you're the one reviewing everything, your throughput caps out at your own attention span. Treat the pipeline as a strong starting template, not a proven machine.

Conclusion

The takeaway isn't "replace your team with agents." It's that one person with a narrow, well-scored agent workflow can now get through the boring front half of a project before breakfast — idea grilling, competitor research, four design options, a bug fix waiting for approval. The humans keep the parts that need a pulse: taste, relationships, and telling the machine precisely where it went wrong.

One action for today: pick the single most repetitive task on your plate this week, write a prompt that asks for a numbered list of ten specific findings with quotes, run it once for free at mykreatool.com, and score the result out of 10. If it lands under 8, rewrite the prompt. If it lands above, you've just found your first agent job.