TL;DR: Method gives your coding agent a saved workflow to run, test, and improve, so the next attempt can build on what worked and fix what didn’t. The Method format and executor are free and open source forever. Use a .method file like you would a SKILL.md or an AGENTS.md file.
I spent the past year productizing agents and it has been frustrating.
I built companies around vibe-coding agents, social media research agents, and data science agents. I ran into consistent problems prompting Codex to build automations.
I’d ask it to fix the problem. It would fix the output without updating the workflow, or add another instruction without removing the conflicting one.
It is likely a familiar story.
What actually is a Method?
A Method is a YAML file with steps, inputs, outputs, and checks. It ships with an open-source executor that runs those steps, manages Codex/Claude Code agents, and records what happened.
You work in your existing Codex or Claude Code setup. Tell your agent:
Turn this task into a Method. Here’s an example of a good result.
Methods are easy to iterate on because the Method is a tangible thing you can see and comprehend. You tell Claude:
I don’t like the result of that run because <3 reasons the result sucks>. Please fix the Method so that doesn’t happen, then test it against the same example.
Methods are versioned, so you can split, test, check, and quantify as you iterate towards a goal. Method is bitter lesson pilled.
I put this to the test to see if Methods help Codex hill-climb on strategies to play the video game Factorio. An Astra-built factory-repair Method used 56% fewer input tokens across 20 test cases and had a slightly lower error rate than using Codex. OpenAI gave me $15,000 in credits for the work. See more.
Get it working once on your laptop
The natural flow for me is: I do the work once, turn it into an automation, deploy it, then give it to customers.
So that’s how Method works. Tell your agent: “Hey deploy this.”
Method packages the workflow with the dependencies and browser accounts it needs. It vacuums all of those into a package, and you deploy it to your cloud.
My ask
Introductions to founders, CTOs, or implementation leads at vertical AI companies building agents for customers. [email protected].
I’m scaling Method, so I’m opening 10 spots to work with teams.
We'll choose one or more of these as the outcome. If these sound like your KPIs, reach out:
- Double the number of agents per day your implementation team builds
- Halve the amount of time one customer-facing agent takes to complete a task.
- Halve the number of tokens per successful run on a customer-facing agent

