Coding Agents Can Rebuild Any App — But Read Nothing First
Coding agents are proving they can rebuild old and new apps alike, but new wire-level analysis shows massive hidden token overhead before they even see your prompt.
The signal: A viral HN thread showing coding agents rebuilding decades-old and brand-new apps alike is trending today, right alongside two separate posts dissecting exactly how much token overhead those same agents burn before they even process your first prompt.
Why it matters: Capability demos make headlines, but the plumbing underneath decides your actual bill and latency. If your agent sends 33,000 tokens of system scaffolding before it reads what you typed, that’s real money and real wait time on every single call, multiplied across a team. Builders evaluating coding agents right now are picking tools based on demo videos, not on what’s actually happening over the wire — and that’s a mistake you’ll pay for at scale.
Are coding agents actually efficient, or just impressive?
No — they’re impressive and bloated at the same time, and those are two different axes people keep conflating. The rebuild demo shows real capability: agents can port and modernize apps across eras with minimal hand-holding. But the token-overhead analysis shows Claude Code sending roughly 33k tokens before it even touches your prompt, versus OpenCode at around 7k — a 4-5x difference in baseline cost for the exact same task. Separately, someone actually reverse-engineered what xAI’s Grok CLI sends over the wire, which is the kind of scrutiny every agent vendor should get and mostly doesn’t. Capability and efficiency are being marketed as one story when they’re two, and only one of them shows up on your invoice.
The pattern I’m watching: We’re entering the phase where agent tooling gets audited the way we used to audit npm dependencies — not for what they can do, but for what they’re quietly shipping in the background. Pair that with the research showing AI narrows the range of ideas people explore even as it speeds them up, and today’s real theme isn’t “agents are amazing,” it’s “agents have hidden costs and hidden defaults you didn’t choose.” The “I love LLMs, I hate hype” post landing today isn’t a coincidence — practitioners are separating the tool from the marketing faster than the market is.
What I’d do with this: Before you standardize on a coding agent for your team, inspect its wire traffic yourself — don’t take a vendor’s token-efficiency claims on faith, especially after seeing what the Grok CLI analysis uncovered. If you’re running agents at any real volume, benchmark system-prompt overhead across two or three tools with the same task, because a 4x difference in baseline tokens compounds fast on a monthly bill. And treat the “can rebuild any app” demos as a ceiling on capability, not a floor on cost.
Key takeaways
- Coding agents can now credibly rebuild both old and new applications, proving real capability gains beyond simple autocomplete.
- Claude Code sends roughly 33,000 tokens of overhead before reading your prompt, while OpenCode sends closer to 7,000, a meaningful cost and latency gap most teams haven’t measured.
- Wire-level audits of tools like xAI’s Grok CLI are becoming necessary due diligence, not optional curiosity, for anyone building on top of these agents.
- Capability demos and efficiency data are two separate stories, and builders who conflate them will overpay without realizing it.