Learn the system
Ben's Bites
A newsletter writer argues real coding skill now means learning the system, not memorizing syntax. It matters because every AI lab is racing to build tools around exactly that idea.
Based on reporting by Ben's Bites — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Ben Tossell's latest newsletter starts with a confession: he can't code, not really, but he's building things anyway. The trigger was a video titled "Agentic coding is a trap — and we all fell for it," and the takeaway he keeps returning to is simple. Learn the system, not the syntax. That's the line separating vibe coding, where you prompt and pray, from agentic engineering, where you understand how the pieces fit even if you never memorize a for-loop.
He traces this back to 2018, when he was stitching together Webflow, Airtable and Zapier without calling it engineering. It was, in hindsight, the same instinct: learn the architecture, let someone else's tool handle the syntax. Code has replaced the no-code stack now, but the muscle is identical. And the reason it's clicking for him today, he says, is that competent engineers are building skills and scaffolding — AGENTS.md files, setup docs, better models underneath — that paper over years of syntax he never learned.
That framing lands differently against this week's product news. Claude Code now lets you run multiple agents in one terminal window and hop into any background session with a slash command. Codex just started working inside Chrome on both Windows and Mac, quietly operating across tabs instead of hijacking the browser. These aren't flashy demos, they're plumbing. They're the system Tossell is talking about, getting easier to hold in your head because someone else is holding the syntax for you.
OpenAI, meanwhile, is betting the same logic scales to entire companies. It bought a 150-person AI consulting shop called Tomoro and is putting $4 billion behind a new deployment arm that partners with major consulting firms to build agent systems for other businesses. Read alongside Tossell's newsletter, the implication is obvious: OpenAI isn't just teaching individuals to stop worrying about syntax, it wants to sell that same shortcut to entire knowledge-worker teams, with Codex as the tool they graduate into once they've learned the system.
Thinking Machines, in contrast, showed off new "interaction models" with real-time voice and video input this week but won't let anyone actually touch them. History with Advanced Voice Mode and Gemini Live suggests these things look magical in a controlled demo and fall apart the moment real users get their hands on them. Which is maybe the whole point of Tossell's essay: the system matters more than the flashiest capability, and you don't really know a system until you've built inside it, mistakes and all.
My take — AI-written commentary, not fact-checked reporting
I like the honesty here more than most 'I learned to code with AI' posts, because Tossell admits he's not aiming to be a real engineer and doesn't pretend otherwise. But the OpenAI-Tomoro move is the story people are sleeping on this week: a $4 billion bet that companies will pay to outsource the system-learning entirely, which quietly locks whole workforces into Codex the way Zapier locked people into its own ecosystem back in 2018. Withholding Thinking Machines' new models from actual users while claiming they're impressive is the same old trick, and I'll believe the hype once real people get to break them.
Read more about this at: Ben's Bites