@ray_fu — 8x breakout

@ray_fu — 8x breakout | Overshot

On-screen text

Loop engineering for Beginners everyone's talking about Code contributed per person, by quarter App.tsx <X> My Wordpress site running on Nginx and Ubuntu 20.04 is experiencing High TTFB and 502 Bad Gateway. If the error "Upstream sent too big header" provides configuration options, then it is likely that the Nginx server is either "time-out" or provide configuration options for execution time. I'll help you troubleshoot these issues. Let's provide solutions for both scenarios: For "Upstream sent too big header" issues: This occurs when PHP-FPM sends headers that exceed the Nginx default buffer sizes. Add these directives to your Nginx configuration: fastcgi_buffers 8 16k; fastcgi_buffer_size 16k; Now for one build that you can right now: Coding agent Add retry logic to the charged call Your time spent waiting / typing 00:25 THE BOTTLENECK What a loop actually does AUTOMATIONS Find work Hand it to the agent Check result Record what happened Decide next STATE FILE The agent forgets each run. The file does not. NEXT MOVE YOU DESIGN IT ONCE then prompts the agent Find work 81 Hand to agent Check w/ test 83 Record it 84 Decide next 85 Is this task ready for a loop? Repeats weekly The task has to come back at least once a week, or the loop never pays for itself. W1 W2 W3 W4 W5 W6 Automated check A check that can fail bad work on its own -- with you out of the room. GATE: FAIL State file So a run resumes where the last one stopped instead of starting over. { "last_step": "Classify", "cursor": 42, "status": "resume" } state.json resuming from step 42 Hard stop A ceiling so the loop can't burn tokens forever if something goes sideways. token budget cap: 58k / run A loop is 5 pieces wired together Automation Fires the loop on a schedule or an event. cron webhook Skill file Project context so the agent never starts from zero. SKILL.md State file Records each run so the next one resumes, not restarts. state.json MCP connectors Open PRs, update tickets, ping you on Slack. github jira slack Sub-agents Split apart -- a model grading its own work always thinks it's right. Writer drafts the fix Checker grades it Writer (model) Done -- added the retry logic. Checker (same model) Looks correct to me. Writer (model) Great, shipping it. -> no real test ever ran Scanning last night's test failures 5 2 1 CLAUSIFIED AS NEEDS HUMAN rounding logic FLAKY network stub timing EASY FIX off-by-one in TTL LOOP ACTION summarize re-run opened PR #308 Two ways a loop quietly goes wrong BATCH OUT No real gate configured Nothing can fail the work. The model that wrote the code is the same one grading it -- so it always passes. SELF APPROVED 1x and counting WHAT'S ACTUALLY TRUE Read tests run Off-by-one bug still present Tokens spent 2.8k -> no real test ever ran -> Read tests, seems right to me. Approved. -> point changes amount code shipped the gap what you understand 10 step road map to learning Loop Engineering

Transcript

Everyone's talking about loop engineering, but most people don't understand why and how it works. Developers are using it to ship features eight times faster. It's not because they can code more. They stop prompting their AI and are using loop system in order to prompt for them. Here's how it works. Right now when you use a coding agent, you type in the prompt and then read the output. Then you type the next prompt. You're basically waiting at every step in our huge bottleneck. Loop engineering means you build a system that finds the work, hands it to an agent, checks the result with the real test, records what happened, and decides the next move on its own. You design it once and the system prompts the agent from that point forward. But you need four things in place before you build one. Every box has to be green before a loop is worth building. The task has to repeat at least weekly. You need an automated check that can fail bad work without you being in the room. You need a state file so that a run resumes where the last one stopped instead of starting over. And you need a hard stop so that the loop does not burn tokens forever. Then the loop system is built with five pieces: an automation that fires the loop from a schedule or event, a skill file that stores the project context so the agent does not start from zero at every cycle, a state file that records what the loop did so that the next run picks off where the last one left off, connectors to MCP that let the loop open PRs, update tickets and ping you on Slack, and sub-agents that split the writing from the checking because a model that grades its own work will always think that it's correct. You can start today with building a simple loop. CI failure triage is the best first loop every night. It scans your test failures, classifies each one, drafts fixes for the easy ones and escalates the rest to you. Two traps that burnt most people. A loop without a real gate is just an agent agreeing with itself on repeat, and the faster the loop ships code that you did not write, the bigger the gap between what your repo has and what you actually understand. I created a complete 10 step roadmap to learning Loop Engineering. If you want it, just like the video and comment step and I'll send it to you. Follow for more advice on making money and running your business with tech and AI.

Loop engineering for Beginners

Promoted product

Why it worked

Summary

Structure

Product placement

Original caption

More from @ray_fu

Call to action

On-screen text