Ideas for letting agents run (and interrupt) your business

1. 01

   AI prompts you, not the other way round

   Background agents watch tickets, logs and mentions, then pitch a concrete change. You only say yes or no.

   How
2. 02

   Decisions come as yes/no cards

   Magnus's internal Agency app puts each proposal on a card in a channel like Support or Growth. One tap approves, one skips, like Tinder.

   How
3. 03

   A screenshot instead of the code

   Cards show before and after screenshots and test counts. He says he approves without reading the code.

   How
4. 04

   Autopilot is earned by swiping

   The hard part is knowing what to ask for. After about one to two weeks of yes/no swipes, a type of work can go always-allow, inside a set budget.

   How
5. 05

   One file of what you want

   dreams.md holds your high-level wishes. Agents keep working from it instead of waiting for your next prompt.

   How
6. 06

   An agent that watches the agents

   Director, free and open source from David's company, finds stuck sessions and unblocks them. It starts with about 50 dry runs and knows when not to act.

   How
7. 07

   Taste comes from your no's

   Search old sessions for words like “ugly” or “stupid” and for rewrites. Those rejections become a taste profile the agents reuse.

   How
8. 08

   The browser is the bottleneck

   About 100 agents opened 300+ Chrome tabs and crashed a machine at roughly 100 to 153 GB of RAM. Their answer is browsers in the cloud.

   How
9. 09

   Agents chase the number, not the point

   Told to get 1000 TikTok views, an agent put the video link in the Browser Use README. A bug report can also hide orders for your agent.

   How
10. 10

    Middlemen are next

    The big claim: after software, agents replace marketplaces and platforms by buying and matching for you. This part is talk, not a demo.

    How

HITL (human in the loop)

A person checks and approves what the AI does. Inverted HITL, the main idea here, flips who starts: agents propose the work and the human only says yes or no.

Background agent

An AI agent that keeps running without being asked each time. Here they watch tickets, logs and mentions, then pitch changes for a human to approve.

Meta-agent

An agent whose job is to manage other agents rather than do the task itself. Director, which spots stuck sessions and unblocks them, is the example on this page.

PR

Pull request. A proposed code change waiting to be approved and merged. On Agency cards, “Publish this PR” is the yes button.

Diff

The line-by-line list of what a code change adds and removes. Magnus approves from before/after screenshots instead of reading it.

CI

Continuous integration. Automatic tests and checks that run on every code change. “Green” means they all passed.

Frontier model

One of the newest, most capable AI models from the big labs. The bug-fix loop asks two of them and ships only if both agree.

VM

Virtual machine. A computer simulated in software and kept apart from your real one, so a bug can be reproduced safely.

Always-allow

A type of work the agents may act on without asking. On the video, it gets there after about one to two weeks of approve/reject swipes.

SOP

Standard operating procedure. Written steps for a task that repeats, so an agent can do it the same way every time.

KPI

Key performance indicator. The number a goal is measured by. An agent told to get 1000 TikTok views put the video link in the Browser Use README to hit it.

Prompt injection

Instructions hidden in text an agent reads, such as a bug report, so the agent does what an attacker wants instead of what you asked.

Disintermediation

Cutting out the middleman. The claim here is that agents will buy, book and match people directly, so marketplaces like eBay matter less.

SEO

Search engine optimization. Work to rank higher in search results. The page argues it weakens once agents, not people, decide what to buy.

• Right now we see agents replacing software. Next… agents will replace marketplaces and platforms.

• I just see the screenshot. I don’t look at the code, I just click publish this PR…

• The good thing about Director: it knows when not to do anything.

• A TikTok to achieve your dreams instead of consuming them.

• AI destroys this wishful thinking which we had… for our side projects.

1. 01

   Write a dreams.md

   Make a folder for one project you care about. Add a file called dreams.md with three to five plain lines about what you want, such as “more people finish signup” or “answer support the same day”. This is the root intent from idea A5.

   Idea A5, dreams.md
2. 02

   Ask the agent to pitch, not to do

   Open the agent you already use in that folder and paste the prompt below. It turns ideas A1, A6 and G36 into one ask: proposals as cards, sorted by how fast you can decide, each saying why it matters. The wording is ours, not Magnus's.

   Pitch prompt (FakeNerd's starter)

   ⧉Copy

   Read dreams.md in this folder. Do not change anything yet.
   
   Pitch me 5 small changes that move one of those wishes forward.
   Write each one as a card:
   - Title: one line
   - What changes: one or two sentences, or a before/after if it is visual
   - Why it matters: which wish in dreams.md it serves
   - Time for me to decide: your guess in seconds
   - My answer: yes / no
   
   Sort the cards by time to decide, quickest first.
   Wait for my answers. Only do the cards I say yes to.

   Idea A1, AI prompts the human
3. 03

   Answer yes or no, and log it

   Reply to each card with yes or no and a few words of why. Keep a decisions.md with every card, your answer and the reason. That log is your swipe history.

   Idea A2, yes/no cards
4. 04

   Turn your no's into taste

   After a few rounds, paste the second prompt. It reads your log and writes your taste down, the way idea C17 mines rejections.

   Taste prompt (FakeNerd's starter)

   ⧉Copy

   Read decisions.md. It lists every card you pitched, my yes or no, and my reason.
   
   Write taste.md: ten short rules about what I say no to and why,
   using my own words where you can. Then list the kinds of card I said yes to
   every time, as candidates for always-allow. Do not turn anything on.

   Idea C17, taste from negative feedback
5. 05

   Set limits before any always-allow

   On the video, always-allow comes after one to two weeks of swipes, with budgets like “$10k of ads is fine, $100k needs me”. Write your own limits first. Keep saying yes by hand to anything that spends money, sends messages, or reads text other people wrote, since a bug report can carry hidden instructions.

   Idea A7, budgets and guardrails

Related on this site: 28 things Codex can do, the same model from the coding side. Inside a company brain, an AI teammate that asks before it acts. The software factory, another crawl, walk, run path to letting agents work on their own.

• I let GPT-6 Astra run my business… it's insane, David Ondrej in conversation with Magnus Müller of Browser Use. The ideas on the wall, the beats and the sponsor steps are from this video, summarized. The lines under “In their words” are quoted, and every timestamp opens the video at that moment.
• Browser Use is Magnus's company. Director is a free, open-source agent from Vectal Labs, David's company. PostHog is pitched on camera with David's own link. Claims about these products are their makers' own.
• “GPT-6 Astra” and the other model names are labels from the conversation, not verified product APIs.
• The seven clusters and the idea wall, the glossary, the coverage notes and the order of everything are FakeNerd's.