An agent harness that is 300x faster and 600x cheaper: cheat code or marketing hype?
A PDF claims it can make coding agents 300x faster and 600x cheaper. Is it really that miraculous? Some people in the comments are already questioning it.
Diogo Almeida, founder of jev, released a PDF explaining how to build a jev harness. In the circles running coding agents, this PDF is being treated as a cheat code.
Promise: up to 300x faster and up to 600x lower cost. Both the speed and savings come entirely from the harness layer outside the model.
Usage is simple too: feed the PDF and Iron Giant's article into Claude Code or Codex and have it rebuild its own configuration following them. Supposedly, in 2 hours, your agent can be running on a harness that few teams have.
What is a harness? In his article, Iron Giant gives a formula: agent = model + harness. The model is Grok or Claude; the harness is everything outside the model: the loop that runs it, the tools it can touch, where memory is stored, what it must not do, and how work is checked. With the same model and different harnesses, you get different agents.
Iron Giant gave an example: Cursor once gave an agent swarm an 835-page SQLite manual and had it rewrite SQLite in Rust. With the same model, the old harness and the new harness reduced merge conflicts from more than 70,000 to fewer than 1,000.
Iron Giant breaks a runnable agent into four layers: Grok writes, jev decides, the harness executes, and the swarm scales. It is specifically split into six parts: chief (routes but does not do the work), contract (each bot's responsibilities, sources, decision criteria, outputs, and forbidden zones), /workspace (uses files instead of chat so context is not lost), the three gates (source, evidence, action), HANDOFF.md (must be written at the end of every task), and ACTION_LOG.md (an audit log).
Most critical step: move decision-making out of the generative model. Most steps in an agent loop do not write anything; they make decisions: which file to open, which model to use, whether it can run, whether it is finished. These decisions used to be handed to the same expensive model and answered with long passages of text. Now they can be handed to jev, a model designed specifically for decisions, costing only $0.042 per million input tokens, with free output and 70–500 ms latency. For one routing decision, Grok 4.7 costs about $0.0049, while jev costs only $0.000084, a 58x difference.
But where does that "300x" come from? Someone in the comments directly asked: "How was the 300x figure measured, and what baseline was it compared against?" No one has given an answer so far. The benchmark details in the PDF are not public, and Iron Giant's article does not mention them either. A number like this is most likely a result from a specific scenario, with limited reference value. What is truly worth paying attention to is the idea itself of "letting a cheap model make decisions."
If you want to try it, just feed the PDF and article to Claude Code or Codex. But do not expect a complete transformation in 2 hours. First get chief, contract, /workspace, and the three gates set up, then consider adding bots. Whether it is 300x or not does not matter; the tokens saved are real.
发布时间: 2026-11-23 02:56