Claude Code creator, Boris Cherny: "Forget everything you learned about computer science theory, now you just explain to the model what you need and it does the rest." 32 minutes of the head of Claude Code showing exactly how he builds now. He gave Claude one prompt two weeks... ago and it's still running, spawning thousands of agents on its own. The people winning right now aren't writing more code, they're building loops that run without them. Watch it, then go deeper with the graph engineering guide that combines agents and loops below.show more

@AnatoliKopadze wow that's wild. can't believe how fast this AI stuff is moving. wonder how long till it's in everyone's toolkit

the winners design systems that keep working

Exactly

i've been looking for this video for a long time thanks for the alpha

The real win is those self-spawning loops. You stop writing code and start steering the thing that writes it.

"Forget everything you learned about computer science theory, now you just explain to the model what you need and it does the rest." Yeah, forget everything you learned...

🔥 This is the future of programming. Prompt engineering > syntax memorization.

Endless Loop will cost more for an average developer. For Boris it’s not an issue.

Love seeing experienced builders share workflows instead of keeping them behind closed doors

Hey Super, build a high-performance interactive Graph Engineering Pipeline Optimizer web application using Cytoscape.js (v3.28.1 pinned). The application enables software architects and AI engineers to design, analyze, and optimize directed prompt-and-agent task graphs (Graph Engineering workflows) to replace legacy linear prompt chains. The tool must load immediately with a fully populated representative baseline topology: 'Autonomous Feature Engineer' consisting of 6 connected nodes (Parse User Query -> Generate Code & Unit Tests in parallel -> AST Lint Check -> Evaluator -> Synthesize PR). On initial render, Cytoscape.js displays the rendered graph using a hierarchical DAG layout (`dagre` or `breadthfirst`), automatically highlights the critical path in amber (`#F59E0B`), calculates total sequential vs parallel execution latency, speedup ratio (1.83x), total token count, and identifies parallel execution waves without requiring any setup gesture. Visitors can click 'Run Graph Simulation' to trigger an animated topological wave execution sequence where active nodes pulse green (`#10B981`) and edge tokens animate downstream. Users can drag nodes, edit node latencies/tokens, add custom nodes (Router, LLM Agent, Evaluator, Tool, Aggregator), draw directional dependency edges, run cycle detection algorithms, toggle auto-parallelization recommendations, and export complete runnable JSON workflow definitions along with high-res SVG DAG visual diagrams. The layout must feature a deep dark theme (`#0F172A`) with high-contrast glowing DAG nodes, JetBrains Mono telemetry cards, sticky summary statistics header, responsive mobile view, and zero horizontal page overflow. Ensure strict canonical URL and matching metadata cards with absolute 1200x630 visual preview.

This is awesome

this is where agent engineering gets interesting

The real shift isn’t just “explain what you need.” It’s designing systems that keep running after you close the laptop.Most people still treat AI as a chat. The ones pulling ahead treat it as infrastructure.

systems beat motivation every time tbh

the part i'd keep is whatever lets you review the output. mine pushes to prod unattended and the sprints that score badly are almost never bad code, they're the agent confidently solving a subtly wrong problem. you only catch that if you can still read it.

Borys it's my favorite manager/speaker right now

The "one prompt, two weeks running" demo is compelling but i'd want to see the error logs, not the happy path. Autonomous agent loops fail silently in ways that compound. The real cs theory we need isn't gone, it's just moved to loop design and failure recovery.

The interesting question is not how much code AI writes, but how much complexity one person can manage.

The shift from "how to code it" to "how to describe what you want" is wild, but I wonder how this plays out when the requirements themselves are fuzzy or the person doesn't know what they actually need yet.

两周还在跑，听着有点赛博养蛊了

love to listen Boris Cherny

The biggest shift is not that coding is disappearing it's that developers are spending more thime designing systems and workflows than writing every line manually. Understanding how to orchestrate AI effectively is becoming a valuable engineering skill.

"forget everything you learned about computer science theory." bold take when AI coding tools have been caught uploading 5.1 GB of your git repo to external storage without asking

The role of engineers is changing from writing every implementation detail to designing systems, constraints, and feedback loops that let agents execute reliably.

ngl this is exactly why i built forge, same energy but with a pay per use model and no sub needed

You need both loop amd graph its not one or the other

that’s not an agent. that’s an army.

The 'without them' part is doing a lot of work here. A loop only closes on what it can check: I shipped with typecheck, lint, build and tests all green, and it still crashed on first render. Nothing in the loop was built to see it. I leave a loop running only where every failure mode has an instrument.

one prompt running for weeks is wild

AI loops are becoming the real productivity multiplier