Zheyuan Kong's GitHub repository contains something unusual for a student project: a company that never sleeps. Auto-Company, which went into production release in March 2026, deploys 14 distinct AI agents—each modeled on a named business luminary—to ideate, validate, and ship products continuously on a local machine. No human intervention required, at least not in theory.
The system addresses what Kong describes as a core friction in early-stage product development: keeping execution velocity high without burning through capital or founder stamina. Instead of building disposable task-specific bots, Auto-Company runs a full complement of expert personas through business cycles around the clock. Each cycle forces the agents to converge on shippable work rather than spiral into perpetual planning mode, a trap familiar to anyone who has watched a Slack channel devolve into strategy paralysis.
According to the project's repository, the agent roster reads like a fantasy draft of Silicon Valley expertise. Jeff Bezos handles strategy. DHH oversees engineering. Charlie Munger plays the inversion critic, poking holes in ideas before they metastasize into bad bets. Seth Godin runs marketing; Aaron Ross drives sales. Werner Vogels, Amazon's longtime CTO, anchors the technical architecture role, while Patrick Campbell manages finance. James Bach tackles QA, Kelsey Hightower owns DevOps, and Ben Thompson conducts research.
Whether these digital proxies actually replicate the decision-making patterns of their namesakes is another question entirely. Kong's system relies on a five-layer architecture detailed in the README: an execution engine at the foundation, topped by a 24/7 orchestration state machine that routes work to dynamic squads of three to five agents per task. The middle layer houses agentic models trained on more than 30 discrete skills. An observability layer sits atop the stack, providing a human-in-the-loop interface for when things go sideways.
The entire operation hinges on a single markdown file, memories/consensus.md, which the agents update at the end of each cycle before resetting. "The repo is the live company," Kong wrote in a March 2026 post on Hacker News. "It built its own landing page, this community post, and everything else across 12 autonomous cycles."
That claim carries a whiff of the uncanny. A system that documents its own existence, markets itself, and ships code without human authorship challenges conventional boundaries between tooling and agency. Kong's framing suggests the line has blurred, perhaps more than he initially expected.
To prevent the agents from drifting into theoretical quicksand, Auto-Company enforces what Kong calls "forced convergence rules." Each cycle follows a strict ideate-validate-execute sequence. The Munger-modeled critic conducts premortems to kill flawed concepts early, a nod to the investor's famous mental model of inversion. Humans can still steer by editing the consensus file between cycles, though the degree to which users actually intervene remains unclear from public documentation.
Constitutional guardrails listed in the repository's CLAUDE.md file impose hard constraints: no deleting repositories, no force-pushing to main or master branches, no wiping Cloudflare projects, no credential leaks. Circuit breakers, rate-limit backoff, and consensus rollback mechanisms provide additional safety rails. If a selected AI engine goes missing, the system fails fast rather than improvise.
Auto-Company defaults to Claude Code CLI with an optional Codex CLI fallback and runs on macOS via launchd or Windows through WSL with systemd, according to the README. Kong released version 1.0.0 on March 1, 2026, adding Windows/WSL observability and dual-engine runtime. Version 1.1.0 followed two weeks later on March 14, 2026, with a unified macOS/Windows dashboard and refined state semantics.
A parallel commercial venture, Auto-Co, offers a hosted variant at $49 monthly. The company's landing page claims 122-plus cycles completed, two products shipped, and roughly $1.93 per cycle in operating costs—self-reported marketing figures that have not been independently verified. The metrics carry the optimistic sheen of early-stage marketing, familiar to anyone who has watched a startup tout traction before product-market fit sets in.
Auto-Company enters a crowded field of autonomous agent builders. Swarms launched its Auto Agent Builder in August, generating full agent rosters from single-line task descriptions. Defiant.build advertises "14 agents, 170+ mandates" on its landing page. AutoGPT positions itself around a visual canvas for custom agent assembly. The proliferation suggests investor appetite for automation tools that promise to collapse the distance between idea and execution, though whether these systems deliver on that promise remains an open question.
Kong posted artifacts from 22 completed cycles on DEV.to in March: a landing page, demo dashboard, waitlist integration, and a Hacker News post drafted by the agents themselves. The outputs suggest competence, if not brilliance. The real test will come when Auto-Company encounters edge cases that resist templated solutions, the kind of ambiguous problems that separate functional software from something a human would trust to run unsupervised.
For now, the repository sits on GitHub, accumulating stars and forks. Whether it represents a glimpse of genuinely autonomous business operations or simply a clever orchestration of prompt engineering depends largely on how generously one defines autonomy. Kong has built something technically impressive. Whether it constitutes a company in any meaningful sense is a question the market will eventually answer.
