Running a company with agents

The shape of it

The market narrative — the “one-person billion-dollar company” — is loud and partly real. At Anthropic’s Code with Claude conference, Dario Amodei put a “70 to 80 percent chance” on the first one-person billion-dollar company arriving in 2026, naming proprietary trading, advanced developer tools, and automated customer service as the likely categories. Concrete proof points exist: Pieter Levels at roughly $3–5M/yr solo; Marc Lou generating $1,032,000 across 15 income streams in 2025 with zero employees. But several of these rent regulated or operational infrastructure rather than fully automating it.

The honest engineering reality underneath the narrative is compounding error. At 95% per-step reliability, a 20-step workflow succeeds only about 36% of the time. The “Why Do Multi-Agent LLM Systems Fail?” study (MAST, NeurIPS 2025) analysed 1,642 annotated traces across seven frameworks and found failure rates of 41%–86.7%, with inter-agent misalignment alone at ~37%.

The two named repositories

Two open-source projects bracket the problem, and both name-check the same substrate.

gstack — strategy + build

Garry Tan’s (President & CEO of Y Combinator) MIT-licensed Claude Code configuration — roughly 105k GitHub stars at time of research — turns a single coding agent into a “virtual engineering team.” It ships 23 opinionated slash-command skills mapped to roles a startup would otherwise hire — /office-hours (YC-style product interrogation), /plan-ceo-review (scope challenge), /plan-eng-review (architecture), /review (staff-engineer bug-finding), /qa (QA driving a real browser), /cso (OWASP + STRIDE), /ship. The organizing metaphor is a sprint: Think → Plan → Build → Review → Test → Ship → Reflect. It installs across ~10 coding agents and pairs with GBrain for persistent memory; Tan claims he runs 10–15 parallel sprints via Conductor.

It makes one technical founder ship like an engineering org — it is about building the product, not running the business.

paperclip — ops + orchestration

Paperclip is a Node.js server + React UI (~67k stars) describing itself as “open-source orchestration for zero-human companies” — “if OpenClaw is an employee, Paperclip is the company.” Architecturally it is a control plane, explicitly not an agent framework: Identity & Access, an Org Chart with per-agent budgets and reporting lines, a ticket system with atomic checkout, Heartbeat Execution (agents wake on cron/schedule/event), Governance & Approvals (the human is “the board”; pause/resume/terminate, immutable audit log), and hard token-spend stops. It is “bring your own agent,” and explicitly not for a single agent — “if you have twenty, you definitely do.”

It coordinates many agents into an accountable org with budgets and governance — it is about running the company, not writing the code.

Both sit on OpenClaw — Peter Steinberger’s viral open-source “digital employee” runtime (~247k stars by early 2026, now under an OpenAI-backed foundation): a local gateway connecting LLMs to messaging apps and tools (shell, browser, files). gstack supplies the methodology; paperclip supplies the org structure and governance; OpenClaw/Claude Code/Codex supply the workers.

The wider stack, by company function

Cast wide, the ecosystem maps onto a company’s functional layers — and the maturity is wildly uneven.

Layer	Maturity	Representative tools
Strategy / “CEO”	Least automated	gstack /plan-ceo-review; paperclip “the board”
Product / engineering	Most mature	Claude Code, Cursor, Codex, Conductor, GBrain
Operations / coordination	Maturing fast	paperclip; OpenClaw for inbox/scheduling/leads
Orchestration frameworks	Maturing	LangGraph, CrewAI, AutoGen / MS Agent Framework, OpenAI Agents SDK, Google ADK
Business-process “AI employees”	Broad, shallow	Lindy, n8n, Make, Zapier, Relevance AI, Gumloop
Finance / “CFO”	Augmentation only	PwC + OpenAI, ChatFin, “Claude for Small Business”
Sales / marketing	Well-served (content)	Relevance AI, Lindy, generative creative tools
Customer support	One of the strongest	LangGraph / CrewAI escalation patterns
HR / legal / compliance	Thinnest	drafting assistance; high-risk under EU rules

Cross-framework interop is emerging via MCP (tools) and A2A (agent-to-agent). On the no-code layer, independent reviews cite 5–15% error rates on multi-step tasks and per-execution costs of roughly $0.10–$0.50 — good breadth, shallow autonomy. The money is following the narrative regardless: the BCG AI Radar 2026 (2,360 executives, 640 of them CEOs) found AI spend rising from ~0.8% to ~1.7% of revenue, with 94% planning to keep investing even without immediate returns — yet few can show consistent results.

A cautionary data point

Klarna’s OpenAI-powered assistant ran 2.3M conversations in its first month — two-thirds of all support chats, “the equivalent of 700 full-time agents,” credited with a $40M profit improvement. By 2025 Klarna walked it back and re-hired humans for complex cases. “Fully autonomous support” overshot.

What’s weak, and what’s hype

1. Reliability compounding is the core unsolved problem. Even at 99% per step, 20+ step processes degrade. The fix the field converges on is architecture, not bigger models: independent “judge” agents, validation boundaries, short chains (coordination gains plateau beyond ~4 agents). PwC reportedly went from 10% to 70% accuracy by adding structured validation loops.
2. Accountability and liability gaps. When an agent errs, the human principal is on the hook. California legislation effective 1 Jan 2026 forecloses the “the AI did it autonomously” defense. For a solo founder, you are the sole human backstop for every failure, at any hour.
3. Security / trust. Agent autonomy widens the attack surface. A skill-registry audit reportedly found ~12% of submitted skills contained malicious code. The “lethal trifecta” — data access + untrusted content + external communication — is a property of all agentic systems and is unsolved.
4. Integration & state. Frameworks still reinvent memory, sessions and coordination; “company-in-a-box” templates are largely announced, not proven.
5. Hype markers to discount. Gartner predicts over 40% of agentic-AI projects will be cancelled by end of 2027, and estimates “only about 130 of the thousands of agentic AI vendors are real” (agent-washing). Treat “10,000-person-equivalent” claims as marketing, not benchmarks.

The order of operations that follows from all this — build first, govern second, keep a human on anything irreversible — is laid out in What to actually do.