Notes from the build.
Ideas to innovate, and lessons we've learned while using it to build the platform.
Introducing Proxy: up to 82% fewer tokens than Claude in Chrome
We gave Claude in Chrome and Proxy the same browser jobs on the same model. Proxy read the news on X with 82% fewer tokens, and across six tasks it spent about half as much.
How to run Claude Code and Codex on the same repo
Give each CLI its own workspace on the same repository — its own checkout, its own branch, its own connector — so Claude Code can build while Codex reviews, and neither one ever edits the other's files.
Running coding agents overnight without babysitting
What it actually takes to leave coding agents working while you sleep: a checkout of their own, a reviewer that isn't the builder, a time limit on every run, a way to reach you only when it matters, and a morning readout of exactly what happened.
Hit the Claude Code weekly limit? Split work across Claude and Codex
When the weekly limit lands mid-week, you don't have to wait. Move whole workspaces to Codex with one setting, keep Claude for the work that deserves it, and put the constant background checks on a model running on your own machine.
Managing 4+ Codex / Claude Code sessions without losing track
Four terminals, four agents, and one question you can't answer at a glance: which one is waiting on you? Give each session its own workspace and let the app keep the list, so what you track is what needs you, not where you left each tab.
How to work on 10 things at once (or more)
A roadmap workspace and one Conductor agent can carry ten separate improvements through their own build-review-merge cycle, one after another, so the only thing left for you to do is add the next item to the list.
Build a product team that didn't exist yesterday
Nobody hands you a Product Lead or a roster of Feature Engineers pre-cast. Imagine who you'd actually want running point, then write them into being.
Loop engineering: the real skill isn't the prompt, it's the loop
A single great prompt gets you one good answer. A well-designed loop — with its own trigger, interval, and exit condition — gets you a process that keeps working after you've stopped watching.
Why good ideas get shelved before anyone decides they're bad
Most ideas aren't rejected for being bad — they're rejected on a cost estimate, and the real measure of a new tool is how many of those estimates it quietly proves wrong.
How to know what happened across every project overnight — without opening any of them
A scheduled agent can quietly check every workspace you're running and leave one digest waiting for you, so catching up stops meaning opening ten different projects one at a time.
What comes after the IDE?
An IDE was built around one assumption — one person, at one keyboard, working on one file at a time — and that assumption, not the editor itself, is the first thing worth questioning.
How to save a multi-agent setup so you're not rebuilding it every time
A Strategy captures a set of agent roles once — triggers, prompts, and workspace scope included — so a pattern like Watcher/Addresser or Engineer/Reviewer can be instantiated again with one call instead of rebuilt by hand.
How to stop starting every morning triaging the same GitHub issues
A scheduled agent can check your repo for new issues on its own, then hand each one through a triage-reproduce-fix pipeline of its own threads — so the chore that used to eat the first twenty minutes of your day happens before you've opened your laptop.
Why "save time" is the wrong way to think about AI
Every generation gets tools that make the old work faster — the real question is what work becomes possible that wasn't before.
The assistant that remembers your research
Long research threads lose their best insights to scroll-back. Give the good parts a permanent home.
Watch the feedback, so you don't have to
A Watcher/Addresser pair turns pull request comments into fixes before you've finished your coffee.
A co-founder for your side project
Solo builders rarely have someone to push back on their own work. An Engineer/Reviewer pair fills that gap.
Five ideas, one afternoon
Why exploring a fork in a decision is cheaper than arguing about it — just build both and see.
Domains: giving every agent a memory it can trust
A place for what your codebase has already taught you, so no agent has to learn it twice.
The Conductor pattern: one agent that never writes code
How one agent keeps an entire roadmap moving forward — without ever opening a single file.
Introducing Frontier
Why I built a place for your ideas to keep moving, instead of one more chat window that forgets everything overnight.