July 23, 2026
Less AI chatter, more code chaos
Code mode yields a 99.2% cost reduction in our systems
AI devs slash costs, commenters cheer, cringe, and say “duh”
TLDR: The team says they made their AI dramatically cheaper by having it run a script and return only a summary instead of hauling in piles of raw data. Commenters mostly agreed the idea is smart, but many also roasted the post as overhyped, obvious, and painfully written.
A developer team just dropped a very brag-worthy number: by letting their artificial intelligence write and run a little script instead of dragging mountains of raw data through a chat, they say they cut costs by 99.2%. In plain English, the bot stops reading every receipt in the filing cabinet and gets handed a neat summary instead. That’s the big win — cheaper, faster, less clutter. But in the comments, the real show began: half the crowd reacted like this was a genius breakthrough, while the other half basically yelled, “Yes, obviously!”
The loudest hot take came from people saying this should already be standard practice. One commenter called the old way “serialized natural language,” then argued that using code for repetitive work is just common sense. Another delivered the cleanest slogan of the thread: put deterministic things in code, non-deterministic in LLMs — meaning if a machine can do something the same way every time, stop making the chatbot improvise it. Ouch.
But not everyone was applauding. One brutally blunt reply simply declared, “The article is so painful to read.” Another mocked the post’s tone, joking that this quarter’s buzzwords are “honest” and “vibes.” So yes, the tech got respect — but the writing caught strays. The funniest part? Even critics seemed to agree the core idea is solid: fewer giant data dumps into the chatbot, more scripts doing the boring work. The crowd’s verdict: great trick, dramatic write-up, please stop acting surprised.
Key Points
- •The article compares script-based code execution with raw sequential tool calls and says this pattern had already been implemented in the authors’ swarm before the post.
- •The benchmark task, `workflow-triage`, scans all workflows and cron schedules and classifies automations as dead, failing, or fine.
- •In the measured production setup, the script handled 26 underlying calls internally and returned only one summary object to the agent context.
- •The article reports 25,811 characters reaching context in the script path versus an estimated 3,259,649 characters in the raw path, about a 126x reduction or 99.2%.
- •Using Claude Sonnet 5 input pricing, the article estimates about $0.02 cost for the script path versus about $2.44 for the raw path, while stating raw-path wall-clock and cost were estimated rather than directly run.