Isenberg posts the agent stack for his 50-person company
A long list of tools and prompts, with no cost, latency or error numbers attached to any of it.
Greg Isenberg posted the full stack he says he runs a 50-plus person team across 12 countries on, broken into spec, design, build, user testing, bugs, data, agent tooling and skills. It is a list of named tools and working rules, not a write-up with measurements.
one agent per platform. the x agent never touches email
running a company is now voice notes and inference
What is in it
The front of the pipeline starts with a voice note to a grokbot, then mattpocock/skills /grill-me so the agent interrogates the idea before anything gets built. The spec lands in Linear through the Linear MCP. Exa and Firecrawl read competitor sites, pricing pages and changelogs, Apify pulls G2, Reddit and app store reviews, and Context7 feeds the agent current docs before it touches an API.
Design is Claude Code plus a Paper MCP, four versions of every screen before any code, then picking pieces from two of them and merging. Build runs through pingdotgg/t3code as the control surface, including from his phone, with stablyai/orca running a fleet of agents in parallel, each in its own git worktree. Graphify maps the codebase into a knowledge graph, and Headroom compresses tool outputs before they reach the model, which Isenberg says makes tokens go 20 to 95 percent further.
User testing uses Lightpanda as the headless browser so hundreds of sessions can run at once, with one impatient persona, one confused persona and one on a phone, plus Playwright and the Chrome DevTools MCP for regression. On bugs, the Sentry MCP catches the error and he says the agent reads the stack trace and opens a fix PR in under 10 minutes, with a human approving. A PostHog MCP cron fires at 7am on yesterday's funnel, and any step that drops more than 10 percent triggers the agent to watch 20 session replays and report back.
The operating rules
The habits are the part worth stealing. One agent per platform, so the X agent never touches email. Cheap model for triage, Opus for anything that ships. Every output gets scored 1 to 10 by an LLM judge, and anything under 8 goes back. Every agent writes a five-line log at night so he can read the whole company in two minutes, and every agent mistake becomes a new skill so it happens once.
For people building on it, there are no costs, no latency figures and no error rates anywhere in the thread. This is one founder describing his own tooling, including the under-10-minute bug fix and the token compression range, not a result anyone measured or reproduced.
Know what Jev is good at before you bet on it
One email a day. Under three minutes. What builders found out about Jev yesterday.