Isenberg posts the full agent stack behind his 50 person team

Greg Isenberg has published the full system he uses to run a company of 50+ people across 12 countries, and his opening line is that he talks to his agents more than his human teammates.

GREG ISENBERG

@gregisenberg

X

every agent writes a 5-line log at night so i can read the whole company in 2 minutes

GREG ISENBERG

@gregisenberg

X

every agent mistake becomes a new skill, so it only happens once

The post runs from idea to shipped code. A voice note goes to a grokbot, a grill-me skill interrogates the idea before anything gets built, the spec drops into Linear through the Linear MCP, and an Exa plus Firecrawl agent reads every competitor site, pricing page and changelog. Apify pulls G2, Reddit and app store reviews, with the prompt asking for the 10 complaints that show up most, quoted from users.

Design is Claude Code plus Paper MCP, four versions of every screen before a line of code, then picking pieces from two of them and merging. Build is T3 Code as the control surface, including from his phone, with Orca running a fleet of agents in parallel, each in its own git worktree. Loops run overnight and he checks what shipped over coffee.

The review rules are the real content

The tool list will rot. The operating rules are the part worth stealing. One agent per platform, so the X agent never touches email. A cheap model for triage and Opus for anything that ships. Every output scored 1 to 10 by an LLM judge, and anything under 8 goes back. Every agent mistake becomes a new skill, so it only happens once.

Then the reporting layer. Each agent writes a five line log at night so Isenberg can read the whole company in two minutes. A PostHog MCP cron runs at 7am on yesterday's funnel, and any step that drops more than 10% gets the agent watching 20 session replays to explain why. Stripe MCP pulls every cancellation reason into one doc every Monday.

User testing uses Lightpanda as a headless browser so hundreds can run at once, with one persona impatient, one confused, one on a phone, which is a lighter version of the synthetic user setup he described earlier. Bugs go Sentry MCP to stack trace to a fix PR in under 10 minutes, human just approves.

The humans are still there, and Isenberg is explicit about what they do now. They get version 5 rather than his voice note, and their job is creative feedback, relationships and telling the agents what they got wrong. His closing comparison is the sharpest line in the thread. A human PM reviews the funnel once a week, this reviews it every morning.

The business read is that none of this is a tool purchase. It is a review process with scoring, logging and a feedback loop that turns errors into permanent instructions, and that part works whether you have 50 people or none.