Isenberg posts 49 steps for running a company on agents
The list runs from voice notes filed into a second brain to agents that rewrite the playbooks every Friday.
Greg Isenberg has published a 49 step list for what he calls the most AI pilled company, a full stack that starts with a folder of markdown files and ends with pointing the whole machine at a new niche. It is a longer and more structured version of the stack he described earlier.
feed it data nobody else has. it's the only part a competitor can't download
The models are the same for everyone but your context is the only part that's yours.
The first eleven steps are all memory. A second brain in Obsidian covering how you think, write and decide. Voice notes all day through Wispr, with an agent filing every fact, story and decision. A company brain holding customers, calls, numbers, pricing, playbooks and past decisions with the reasons attached. Calls recorded in Granola with transcripts going straight in. A facts file of true things agents are allowed to say, which Isenberg says kills made-up claims. A voice file with your 50 best posts, 20 you hate and the words you never use. All of it in git, with one instruction file at the top, either CLAUDE.md or AGENTS.md.
One agent per project
The agent section is built around an orchestrator that never turns off, named as Grok Bot, openclaw or hermes. It manages one agent per project, and each of those runs its own sub-agents. Isenberg suggests routing models through OpenRouter so the expensive ones plan and check while cheap ones do volume, tracking cost per job, and killing any job that costs more than it earns.
For the build line he describes a spec agent feeding builder agents running in parallel on git worktrees, then a test agent, a reviewer agent and deploy. The reviewer agents get a named bar, which he writes as asking whether the best team in the world would ship this, and rejecting until the answer is yes. He also recommends keeping 50 real cases with known right answers and rerunning them every time a new model drops so you can switch to the winner.
The loop that compounds
The closing section is the one worth arguing about. Every finished job writes back to the brain. Support tickets, analytics and comments flow in nightly. Every Friday an agent rewrites the playbooks from the week's results and you approve the diff. Every month everything gets ranked by profit per human hour, kill one thing and double one thing.
The business read is in his last line. Everyone has the same models, so the defensible asset is the context you have accumulated, and that is also the part that takes months of filing and recording before it does anything useful. Isenberg frames the humans who are left as the ones who approve, own the outcome and talk to customers, which he puts at 10 percent of the work.
Steal the numbers, not the vibes
One email a day. Under three minutes. Real MRR, real pricing, real launches from founders building with AI.
