ninja nerdingIn loop, near loop, out loop
In loop, near loop, out loop
IndyDevDan argues that the best models now work better inside the loop than we do, so the person typing every prompt is the slow part. His ladder for getting out of the way: in the loop you prompt, near the loop you approve, out the loop you observe. The question he asks about every task: what can run without me next?
Jump to the ladder map, the three agents he runs, or move one task yourself. Want to build the agents? That's on How to use Pi Durable.
The video
26 minutes, posted Oct 26, 2026, on IndyDevDan's channel. Chapters are his. Every timestamp on this page opens the video at that moment. It's the third video in a row on proactive agents; the first two show the on-call and news agents in full.
- 0:00 The Alien Engineer
- 2:02 In Loop, Near Loop, Out Loop
- 5:05 Why Use Proactive Agents
- 8:37 Real Durable Agent Use-Cases
- 14:46 Why Build Durable Agents
- 22:33 Phase 3 Agentic Engineering (his course pitch)
- 23:03 How to Move Out the Loop
The two earlier videos: the on-call agent (Oct 12) and the news agent (Oct 19). Both are taught on the Pi Durable page.
Thirteen signals
The whole video in thirteen boxes. Each links to the part of this page that explains it.
- 010:07
Read the trace
Give a top model a hard task and read every command it ran. Dan says it now works like an alien engineer and makes fewer mistakes inside the loop than you do on a good day.
More - 023:00
So the person in the chair is the slow part
If the agent is better at the work, the human typing every prompt is what holds it back. His claim: 80% or more of today's agent work could run without you, with the right system.
More - 032:25
Prompt, click, cron
The ladder. In the loop you prompt. Near the loop you approve with a click. Out of the loop it runs on a timer or an event and you only watch.
More - 043:41
Step, don't jump
Going straight from typing every prompt to fully automatic is hard. Near the loop is the bridge: the agent does the work, you step in at the one decision that matters.
More - 054:11
It works for you, then waits
A reactive agent waits for you to start it. A proactive one starts on its own when something happens, does the work, and waits for your decision.
More - 064:56
Feed it information, not prompts
What a proactive agent needs to start isn't you. It's the right information: a log line, a ticket, a new file. Getting that to the right agent is the job now.
More - 076:11
The near-loop kit has four parts
An agent that survives crashes, its own computer to run on, a way in from outside (an API), and triggers that call it.
More - 088:43
He runs three of these right now
An on-call agent that fixes errors overnight, a news agent that sorts hype from signal, and a new one that rebuilds a design until the screenshot matches.
More - 0915:35
The core four, times what you know
Context, model, prompt and tools make up every agent. How well you handle those four, multiplied by what you know about software, decides how far out of the loop you get.
More - 1016:22
The harness is where you pull ahead
Everyone has the same Claude Code and Codex out of the box, so the default is no edge. Your own tools, guards and prompts around the model are.
More - 1120:16
Know it so well you don't look
His line: vibe coding is not knowing what your system does and not looking. Agentic engineering is knowing it so well you don't have to look.
More - 1221:00
Count agents, not tokens
His new scoreboard: how many agents are doing useful work for you this second. Not a loop burning tokens for show.
More - 1325:23
Ask it every time you sit down
Is there a reason I'm still in the loop for this work? He says the only good yes is when you're building the system that builds the system.
More
Words on this page
38 terms, in plain words. Skip this if you already run agents on a schedule.
- The loop
- The back and forth with an agent: you ask, it works, you check, you ask again. Who has to be there for each turn is the whole question on this page.
- In the loop
- You type every prompt and watch every answer. Close the laptop and the work stops. Dan's word for it: prompt.
- Near the loop
- Something other than you starts the agent. It does the work and stops at one decision for your yes or no. Dan's word: click.
- Out the loop
- It starts on a timer or an event and finishes without you. You read what it did and improve the system. Dan's word: cron.
- Reactive agent
- One that waits for you to start it. Every normal chat with Claude Code, Codex or Pi.
- Proactive agent
- One that starts on its own when something happens, does the work, then waits for you. On screen: works for you, then waits.
- Trigger
- The thing that starts the agent. His list on screen: cron, webhook, log event, ticket, a command line, a browser, another agent.
- Cron
- A built-in clock on Mac and Linux that runs a command at set times, like every morning at 7. Dan uses the word for fully hands-off work.
- Webhook
- A message one app sends another the moment something happens, such as a new payment or a new bug report. Your agent can listen for it.
- Log stream
- The running list of everything a live app records: each request, each error. The on-call agent watches it.
- Ticket
- A task or bug report in a tracker like Linear, Jira or GitHub Issues. A new one can start an agent.
- API
- A door into a program that other programs can use. Dan's near-loop agents have one, so any trigger, script or app can hand them work.
- Durable agent
- An agent that saves where it is as it goes, so a crash, a restart or a closed laptop doesn't lose the work. Dan builds his on Pi Durable.
- Sandbox
- A separate computer or walled-off space just for the agent: a spare Mac mini or a cloud machine. Always on, and a mistake there can't touch your laptop.
- Jev
- A small, fast, cheap model built for one-word judgments, like normal or incident. Dan puts it in front of his agents as a filter. FakeNerd has a card on it.
- Classify
- Sort something into one of a few labels: hype, facts or signal; normal, suspicious or incident. Jev's job in all three of Dan's agents.
- 4xx and 5xx errors
- Codes a website sends back. 4xx means the request was wrong (404 is page not found). 5xx means the server broke. Dan: if a plain code check is enough, skip Jev.
- Root cause
- The real reason something broke, not just the symptom. Finding it is one step of the on-call agent's job.
- Patch
- A small code change that fixes one thing. The on-call agent prepares it before you wake up.
- AI developer workflow
- Dan's name for a fixed series of steps an engineer used to run by hand, now run by agents: detect, investigate, fix, ask, apply, verify, save.
- Harness
- The program around a model that gives it tools, runs them and feeds back the results. Claude Code, Codex and Pi are harnesses.
- Harness engineering
- Building your own harness, or changing one: adding tools only your job needs, taking away dangerous ones, adding guards. Dan's pick for where the edge is.
- Context engineering
- Choosing what the agent knows when it starts: which files, notes, examples and history go in, and what stays out.
- Tokenomics
- Dan's word for the cost side of picking a model: good enough for the job, not so pricey that an always-on agent breaks the bank.
- System prompt and user prompt
- The system prompt is the standing rules the agent always has. The user prompt is what you, or a trigger, ask each time. Dan says you need to write both.
- Damage control
- A guard that checks a command before the agent runs it and blocks the dangerous ones, like rm -rf, which deletes files for good.
- Hill climbing
- Try, score, change, try again, keeping whatever scores higher. The UI replicator climbs toward a page that matches the design.
- RSI loop
- Short for recursive self-improvement. Dan calls his replicator a primitive one: an agent loop that grades its own work and tries again.
- Match score
- How close the agent's screenshot is to the design you gave it, out of 100. Dan's replicator aims for a line at 85.
- Multi-client
- One agent you can reach from many places: the browser, the terminal, your phone. The agent lives on its machine, not in the window.
- Multiplayer
- More than one person can open the same agent: you, product, design, support.
- Data flywheel
- Each vote you give makes the next pick better, which gets you to vote more. The news agent learns from your thumbs up and down.
- Token maxing
- Running as many agents and tokens as you can. Dan's scoreboard is the opposite: count only agents doing useful work.
- Vibe coding
- Building with an agent without knowing or checking what it made. Dan calls it the floor.
- Agentic engineering
- Building agent systems you understand well enough to trust without watching. Dan calls it the ceiling.
- Meta skill
- A skill that teaches the agent to write skills. Dan's example of you already stepping back from doing the work yourself.
- Orchestration
- One agent directing other agents. Dan's example: instead of running ten agents yourself, you run one that prompts nine.
- Zero-touch engineering
- Dan's name for work that runs with no human touch at all. The goal of his next course.
What it is, and what it does
Not a tool. A ruler for your own work: for each task, who has to be there for it to happen? Then a direction: move each task one rung down at a time. 2:25
- 1 · PromptIn the loopYou type every turn. Work stops when you do.
- 2 · ClickNear the loopIt starts without you and stops once for your yes.
- 3 · CronOut the loopIt starts and finishes on its own. You read the results.
On his slide the arrow underneath is labelled agentic engineering skill. If you're getting better, he says, more of your systems sit near or out of the loop. He warns against jumping straight from the first box to the last. 3:41
The same task on each rung · basic to advanced
A customer hits a bugNear: Dan's on-call agent
In the loop
You paste the error into your agent and go back and forth until it's fixed.
Near the loop
The error starts an agent. It writes the fix overnight. You tap Approve.
Out the loop
Small fixes that pass every test ship on their own. You read a daily list.
Keep up with AI newsNear: Dan's news agent
In the loop
You scroll X and YouTube and get clickbaited.
Near the loop
An agent sweeps on a timer and sorts each story into hype, facts or signal. You vote up or down.
Out the loop
It writes your Monday brief from the signal and files it in your notes.
Turn a design into a pageNear: Dan's UI replicator
In the loop
You prompt, look at the page, prompt again.
Near the loop
Drop the design in. A loop builds, screenshots, scores and fixes until it passes. You review the winner.
Out the loop
Every design added to a folder is built and opened as a pull request.
Send a weekly reportAll FakeNerd's
In the loop
Every Friday you ask the agent to write it.
Near the loop
Friday at 4 it drafts the report and asks you before sending.
Out the loop
It sends itself. You skim the copy in your inbox.
The near-loop column of the first three rows is what Dan runs. Every other cell is FakeNerd's example.
Everything named
Dan teaches the idea, not one product. These are what he names or shows. The one he leans on hardest is the Pi harness, because you can change it.
Models
Anthropic
Claude Opus 5.5
A top model. His opening test: give it a hard task and read the trace.
OpenAI
GPT-6 Astra
A top model. Scored 86.1 in the UI replicator, past the line.
Anthropic
Claude Sonnet 5.5
The cheaper Claude. Reached 81.7 on his landing page in the replicator, then stalled.
DeepSeek
DeepSeek
Named with Opus and Astra as new tools, and again when he says to watch token costs.
TypeSafe
Jev 1.14
The fast, cheap classifier in all three of his agents. The replicator is new since 1.14.
Agents and harnesses
Earendil
Pi coding agent
The open-source harness he recommends building on, because you can add and remove what it can do.
Earendil
Pi Durable
Makes a Pi agent save as it goes, so it survives crashes and restarts. Under all three of his agents.
Anthropic, OpenAI
Claude Code and Codex
Fine in the loop. Out of the box, he says, they're the default everyone has, and the session ends when you close it.
Hosted, named only
GrokBot and Claude Cowork
Hosted products that package always-on agents for everyone. Fine, he says, but this is technology to own yourself.
Where the agents live
Apple
Mac mini (M4 and M6)
Always-on machines at his desk. The news agent has run on the M4 for a week and a half.
Any provider
Cloud sandbox
A rented machine in the cloud. The replicator runs on one, so nothing is local.
Open source
SQLite
The small database file on his slides where a durable agent keeps its state.
Guards
Left out
His Phase 3 course pitch (22:12 to 23:00 and 24:14 to 25:10), and two names the captions garble and the screen doesn't settle: a hosted product he calls “dots”, and the model behind a headline in his news feed.
The ladder map
One piece of work, seven steps. The step names are from his slide. Pick a rung and watch the steps leave your hands, click a step to see who does it on each rung, or press Start to follow one 3 a.m. incident. 5:15
Pick a rung
Near the loop: Something else starts it. The agent works, then waits for your yes at the step that matters. Work happens before and after you check in. Starts with: a trigger. You do 1 of 7 steps. 3:58
One piece of work · shaded boxes are yours on this rung
Follow one 3 a.m. incident, near the loop
Eight beats from Dan's on-call storyboard: errors at 3 a.m., a fix ready by 3:18, your one click at 7. Watch which box lights up, and how little of it is you.
Step 4 of 7
Watch at 5:40Decision needed
The one step that still needs judgment. The agent shows what it found and what it wants to do.
- In the loopYouYou, after watching every step.
- Near the loopYouYou, once. Read the cause and the change, then click Approve or Decline from any screen, even your phone.
- Out the loopYour rulesRules you wrote ahead of time decide. Say: the tests pass and the change is small, so it ships. Anything else still asks you.
What moves it off your plate
A saved wait. The question sits in storage, so a restart doesn't lose it or ask twice.
Watch out
Out the loop is only as safe as the checks behind your rules. Dan's own demos stop here and ask.
Dan shows the near-loop row on one slide and talks through the other two. Which step is whose on the in and out rungs is our wiring of what he says.
The near-loop kit
A normal agent waits for you to work. A proactive one works for you, then waits. On his slide it's a sum. 7:38
Sandbox · its own computer · reached through an API
- 01
A durable agent
Saves as it goes, so it survives crashes, restarts and a closed laptop. It also needs memory: it has to keep track of everything, every time.
On screen: Durable agent · survives crashes, restarts, and closed laptops
- 02
Its own sandbox
Its own computer: a spare Mac mini or a cloud machine. Always on, walled off from your laptop, free to work. The most important tool you can give an agent, he says, is a full computer.
On screen: Agent sandbox · its own compute. Always on. Isolated, safe, autonomous.
- 03
An API
A door into the agent, so anything can hand it work. Not just you typing.
On screen: API
- 04
Triggers
The things that knock on that door. Seven on his slide.
On screen: Triggers · cron · webhook · log event · ticket · cli · browser · agent
Seven triggers from his slide · “why are you still the main one?”7:15
- cronA time: every morning at 7, every five minutes.
- webhookAnother app says something happened: a payment, a form, a bug report.
- log eventA line in your app's live log, sorted by Jev.
- ticketA new task in Linear, Jira or GitHub Issues.
- cliA command you or a script run once and walk away from.
- browserA button in a web page, or a design dragged and dropped.
- agentAnother agent handing off work on its own time.
Why build durable agents · his four reasons14:46
Proactive
It can start before you show up, before you wake up.
Multi-client
Browser, terminal, phone. Anyone, anywhere can connect to the same agent.
Multiplayer
You, product, design and support can all hop into the right agent.
Long-running
It survives crashes and restarts. It has to.
Put together, he says, it removes the seat from under you: you don't have to be in the dev chair. Every screen you use (browser, terminal, phone) is just a window onto an agent that lives on its own machine. 14:08
Three agents he runs
All three sit near the loop: they do the work on their own and stop for you once. Shaded steps are yours. The first two have open-source repos (MIT); the third is new in this video and has no repo yet.
1 · The on-call agent8:43
- 1WatchA live log stream. Jev judges every row: normal, suspicious or incident.
- 2OpenEnough incident labels on one error and code opens an incident and wakes the agent.
- 3FixThe agent investigates, finds the root cause and writes the fix.
- 4ApproveyouYou read it and click deploy. The only step that's yours.
Fixes errors in a live app while you sleep. The demo uses fake logs and a mock shop. Repo: disler/durable-pi.
2 · The news agent10:06
- 1SweepOn a timer, it gathers new stories. His runs on an M4 Mac mini and has been on for a week and a half.
- 2SortJev labels each story hype, facts or signal. Signal is news you can act on.
- 3PresentThe best few land in a feed you open in the browser, written the way you like to read.
- 4VoteyouThumbs up or down. In the video he upvotes a model that scored 40 out of 40 on a prompt-injection test.
- 5LearnYour votes feed the next sweep. The more you vote, the better it gets.
He calls it a product agent: built for one job, it beats the general default by miles. Repo: disler/pi-durable-news-agent.
3 · The Deep Jev UI replicator12:01
- 1Drop a designyouDrag in a screenshot or mock of the page you want. Then walk away.
- 2BuildAn agent builds the page in a cloud sandbox.
- 3ScreenshotIt takes a picture of what it built.
- 4Jev comparesJev scores how close the picture is to the design, out of 100.
- 5Fix and repeatBelow the line (85 on his slide), the agent fixes and goes round again. Each round should score higher.
- 6ReviewyouYou look at the best version. About 30 minutes for all the rounds, he says.
He calls it a primitive self-improvement loop: it grades its own work and climbs. No repo; our starter prompt builds a small one.
Replicator scores, read from the screen · line at 8513:18
GPT-6 Astra86.1
a sample landing page, passed the line
Claude Sonnet 5.581.7
his Phase 3 landing page, best at round 3, then stalled
Two runs on two different pages, so not a fair race between models. What it shows: the loop climbs, and sometimes stops short. He drops the design in, leaves, and comes back about 30 minutes later.
The core four
Every agent is made of four things. His slide: “How far out you get depends on how well you engineer these four.” Then he puts them on top of what you already know. 15:35
Agentic engineering, as he draws it
- ContextContext engineeringWhat it knows
- TokenomicsTokenomicsWhat it costs
- PromptPrompt engineeringThe instructions
- HarnessHarness engineeringWhat it can do
His point for engineers: your years of knowing what to ask and what to build are what the four multiply. For non-engineers, the same goes for whatever you know best about your own work. 18:26
Each of the four, in the loop vs out of it · FakeNerd's table from his points
- ContextContext engineeringIn the loopYou paste in whatever seems relevant, each time.Out the loopThe right information arrives with the trigger: the log rows, the ticket, the design.
- ModelTokenomicsIn the loopWhatever your subscription gives you.Out the loopGood enough, fast enough, cheap enough to run all day. A tiny model like Jev for sorting, a big one for fixing.
- PromptPrompt engineeringIn the loopA user prompt you type now.Out the loopA system prompt you wrote once, plus a user prompt the trigger fills in.
- ToolsHarness engineeringIn the loopEvery tool the default app ships with, bash included.Out the loopOnly the tools this job needs, plus guards on the dangerous ones.
Your own harness
One agent with the default setup is, in his word, saturated: everyone has it. The edge is specializing the harness around the model for your job. 16:22
The news agent's harness, from his slide19:31
Harness · Jev damage control on
Model
- run_news_sweepGo gather the latest stories.
- jev_classifyAsk Jev for a one-word label.
- research_itemsDig into one story.
- ask_tasteAsk you what you like.
- present_newsPut the picks in your feed.
- rm -rfBlocked. Deletes files for good. Jev damage control stops it.
His recipe, said in order19:23
- 1Pick the model on cost, speed and qualityStrong enough for the job, cheap enough to run all day. Subscriptions are heavily subsidized now, he warns, and won't stay that way.
- 2Build your own harnessUse any agent SDK. Dan recommends Pi, because you can add and remove what the agent can do.
- 3Give it tools only your job needsLike the news agent's run_news_sweep or ask_taste. This is where your product gets results the default can't.
- 4Guard the dangerous onesAdd Jev damage control: every bash command goes to Jev first, and destructive ones like rm -rf are blocked.
- 5Or take bash awayIf the agent doesn't need a general command line, remove it and hand it exactly the tools it does need.
On Claude Code instead? The same guard exists as a hook: disler/claude-code-damage-control.
Before and after · “vibe coding is the floor, agentic engineering is the ceiling”19:11
- Do you know what it does?Vibe coding · NoAgentic engineering · Yes, well enough to explain it
- Do you look?Vibe coding · NoAgentic engineering · You don't need to. Checks look for you
- What starts itVibe coding · You, every timeAgentic engineering · A trigger you designed
- What it can doVibe coding · Whatever the default allowsAgentic engineering · Exactly the tools you gave it, with guards
- Dan's wordVibe coding · The floorAgentic engineering · The ceiling
The first two rows and the last are his words. The middle two are ours, from the rest of the video.
How many agents are working for you right now?
His new measure. Not tokens burned, not a loop spinning for show: agents doing useful work for you this second, without you. 21:00
In the loop
As many agents as you can watch. A few terminals before you lose track.
Near the loop
Many agents working, each stopping to ask you once.
Out the loop
A field of agents. You read the results.
The dot counts are a picture, not his numbers. His reason: you can only split your attention across so many terminals before you lose track of what each one is doing and just trust the model. 21:32
Where he says this goes · labels from his slide, glosses ours22:13
- Phase 1 · AugmentAI helps you do the work.
- Phase 2 · AutomateAgents do the work while you drive them. He calls phases 1 and 2 in the loop, and saturated.
- Phase 3 · DeprecateThe work no longer needs you. He calls it zero-touch engineering.
Phase 3 is also the name of the course he's selling. You don't need it to use anything on this page.
Try it, smallest version first
Move one boring task one rung. Steps 1 to 5 cost nothing beyond the agent you already pay for. You can stop after step 5 and you're near the loop.
Which rung for this task? Four questions · FakeNerd's
1Does the task start the same way every time?
Yes
Good. That start can be a trigger: a time, a new file, a webhook, a ticket.
No
Keep it in the loop for now. Agents need a clear way in.
2Is there one decision that really needs you?
Yes
Make that the one click. Everything before and after it is the agent's. That's near the loop.
No
If nothing in it needs judgment and it's safe to repeat, it can go out the loop.
3Can the agent check its own result?
Yes
A test, an error count, a score. Then it can verify and you only review.
No
Keep the check yours, and say so in the prompt.
4Could a mistake cost money, data or trust?
Yes
Near the loop, with your click right before the risky step. Run it on its own machine.
No
Try out the loop, and read the results every day for the first week.
- 01
Sort your week
Paste the first starter prompt into the agent you already use: Claude Code, Codex, Cursor, or a sandbox like Cloudroom or bb (getbb.app). It puts each task you did with agents this week on a rung. Ten minutes, no code.
The starter prompts - 02
Pick one boring task
One that starts the same way every time and has at most one real decision. A morning summary, a report, a folder you tidy. Not production code.
- 03
Run it once by hand, headless
Most coding agents can run one prompt and quit, with no chat window. Try it from the project folder. If the result is good with no follow-up, the task is ready to leave your hands.
Run it once by hand, headless
# Claude Code claude -p "Read notes/inbox.md and write today's three priorities to notes/today.md" # Codex codex exec "Read notes/inbox.md and write today's three priorities to notes/today.md"
- 04
Give it a trigger
A time is the easiest trigger. Many agent apps can now schedule a prompt for you. Or use cron on a Mac or Linux machine: open it with crontab -e and add one line. This one runs at 7 every weekday. Cron has a short search path, so if the log says command not found, swap claude for the full path that which claude prints.
Give it a trigger
0 7 * * 1-5 cd ~/projects/notes && claude -p "Read notes/inbox.md and write today's three priorities to notes/today.md" >> ~/agent-runs.log 2>&1
- 05
Add the one click
Tell it to stop before anything that sends, pays, deletes or deploys and ask you. Use the second starter prompt. You are now near the loop.
Near-loop starter - 06
Move it off your laptop
Your laptop sleeps. A spare Mac mini or a cloud sandbox doesn't. This is the step where you also want the agent to survive a restart: see how on the Pi Durable page.
How to use Pi Durable - 07
Put a cheap filter in front
If the trigger is noisy, like a log or an inbox, let a small model sort it first so the big agent only wakes for the real thing. Dan uses Jev.
The Jev card - 08
Count
Once a week, write down how many agents did useful work for you without you. That's Dan's scoreboard. Then ask: what can run without me next?
Starter prompts
Dan shows no prompts in this video, so all three are FakeNerd's, built on his ladder. Fill in the brackets.
FakeNerd starter · sort my week by rung
Help me see which of my agent work could run without me. Here is what I did with AI agents this week: [LIST EACH TASK IN ONE LINE, e.g. "fixed a bug a customer emailed about", "summarized AI news", "wrote the Friday report"] For each task, tell me: 1. Which rung it is on now: in the loop (I type every prompt), near the loop (something starts it and I approve once), or out the loop (it runs and I only read results). 2. What could start it without me: a time, a webhook, a new file, a ticket, a log line, or "nothing yet". 3. The one decision that really needs me, if any. 4. Whether the agent could check its own result, and how. 5. The next rung it could move to, and the smallest step to get there. End with the single task I should move first, and why. Keep it plain; I'm not an engineer.
Fill-in · add these lines for a sharper answer
Task list: [YOUR TASKS, ONE PER LINE] Tools I have: [e.g. Claude Code on a Mac, a Notion workspace, a GitHub repo] Things I never want automated: [e.g. emails to clients, payments]
Run this one first, read-only. It changes nothing; it gives you the task to move.
FakeNerd starter · move one task near the loop
Turn this task into a near-the-loop agent: it starts on its own, does the work, and waits for me at one step. Task: [WHAT IT DOES, e.g. draft the weekly report from notes/week.md] Trigger: [WHAT STARTS IT, e.g. every Friday at 4 pm, or a new file in ./inbox] The one decision that's mine: [e.g. sending the report] Rules: - Do all the work before that decision on your own. - Then stop. Show me what you made and what you want to do next, and wait for my yes or no. - Never send, pay, delete or deploy without my yes. - Save what you did and why to [FILE OR FOLDER], so I can read it later from anywhere. - If something looks wrong or the input is missing, stop and tell me instead of guessing. First, explain in plain words how you'll set up the trigger on my machine, and what I need to install. Then build it.
FakeNerd starter · a small design replicator
Build me a small design replicator loop. Input: a design image at [PATH, e.g. designs/home.png]. Goal: a page in [STACK, e.g. plain HTML and CSS] that looks like it. Loop: 1. Build or update the page. 2. Take a screenshot of it at the same size as the design. 3. Compare the screenshot to the design and give a score out of 100, with the three biggest differences. 4. If the score is below [LINE, e.g. 85], fix the biggest differences and go back to step 1. 5. Stop at the line, after [MAX ROUNDS, e.g. 6] rounds, or when the score stops going up two rounds in a row. Keep every round's screenshot and score in [FOLDER]. At the end, show me the best round and its score. Work only inside this project folder.
Next to the others
Pages on this site that sit on the same ladder, and where.
- How to use Pi Durable The library under Dan's agents, step by step, with his on-call and news agents to run yourself.How to build near the loop
- Jev The cheap classifier in front of all three of his agents, with prompts and the wiki.The filter that makes triggers useful
- LLM benchmarks Dan's way of picking a model on cost, speed and quality. His tokenomics, in full.One of the core four
- Warp: software factories Crawl, walk, run: another ladder from you doing the work to agents doing it.Same idea, factory words
- Steve Sewell's factory A factory that fixes bugs on a timer.Out the loop, for bugs
- Ondrej's mini software factory Stop starting agents by hand: loops that check their own work and run around the clock.Near to out
- Karpathy's autoresearch An agent left running overnight, keeping only what scores better. The replicator's hill climb, for research.Out the loop, for experiments
- Theo on token-maxing Ten threads at once, all watched by one person. The in-loop ceiling Dan wants you to leave.In the loop, at full speed
FakeNerd's take
The question is the gift
Skip the alien-engineer hype and keep one line: what can run without me next? Ask it about every task and the ladder does the rest.
Near the loop is home for a long time
Out the loop needs checks you trust with your money and your name. Most readers should live near the loop, and only move a task out once it has run clean for weeks.
Start boring, not on production
Dan's examples are an on-call agent and a 24/7 news feed. Yours should be a morning summary or a Friday report. Same ladder, nothing to break.
Always on means always billing
Dan says it himself: subsidized subscriptions won't last. Put a small model in front, cap the rounds, and check the bill in week one.
Attribution
- “ELITE Agentic Engineers Are LEAVING THE LOOP... Here's why” by IndyDevDan (posted Oct 26, 2026). The ladder, prompt-click-cron, the near-loop kit, the four reasons, the core four, the harness recipe, the scoreboard question and the three agents are his, in our words. Slide labels, step names, the replicator scores and the Mac mini model are read from the screen; where the captions disagreed, the screen won.
- He sells courses: Tactical Agentic Coding (his Phase 2) and a Phase 3 course in pre-sign-up, pitched at 22:12 and again at the end. Those parts are left out. In his earlier videos he says he takes no sponsorships.
- Repo facts (the incident rule, the damage-control behaviour) are from disler/durable-pi and disler/pi-durable-news-agent, both MIT, read Oct 28, 2026.
- Not checked: the 80% figure, “agents make fewer mistakes than you”, the 30-minute replicator run, and what other teams are rolling out. Model names are as said.
- The map wiring (who does which step on the in and out rungs), the walkthrough beats past his storyboard, the same-task table, the four questions, the try-it steps, every prompt and command, the sibling list and the takes are FakeNerd's. Not sponsored by anyone on this page.