We tried to runa company on agents.
We were building anAI website agency.
The mission was simple: rebuild small-business websites to $20K premium quality, using AI agents instead of humans. First client on the test bench was Ampha Group, an accounting firm in Sandton, South Africa. We had enrichment data: brand colors, logos, services, testimonials. We had a pipeline: Opus plans → Step Flash designs → Gemini Flash generates images → GPT 5.4 Mini builds → Sonnet QAs. Target cost per site: $25.
It cost us$706 in a day.
We were using Paperclip, a single-task agent runner with a heartbeat system. Every 3 to 5 minutes the agents would reload their full context, which was now 2.4 million tokens deep, and make another LLM call. We burned $706 in 24 hours and the site was not shipped.
“Nothing was working. The designer kept applying the same CSS fix to Header.tsx for two hours. The QA agent kept saying the site looked good. Meanwhile the planner was making pages for services Ampha Group didn't even offer.”
— SESSION LOG, DAY 3Everythingwas broken.
The Designer spent hours applying the same CSS fix over and over, forgetting what it already tried. Child processes were never cleaned up so zombies ate our CPU. Agent crashes left task locks stranded and we had to write raw SQL to unblock work. The QA agent kept approving broken sites. It invented fake offices in London and New York for a company that only has one office, in Sandton.
We tried everything. None of it held. The fundamental architecture was the problem.
So we ripped out the engineand rebuilt it.
Company Agents is that rebuild. Nine parallel research agents surveyed 200+ papers and reference projects. We read every post-mortem we could find. Leases instead of locks, structured checkpoints, loop detection, process-group isolation, and stacked budgets at every level.
And then, just as important, we stopped pretending humans didn't belong. Company Agents is built for hybrid teams. AI ships the work; humans own the decisions that matter.
One agent, two hours, same CSS fix 400 times.
9 parallel agents across orchestration, memory, QA.
~13k lines on top of the forked Paperclip layer.
From post-mortem to v0.1.0 in the wild.
Ran the original Paperclip agency. Watched $706 vanish. Started writing the replacement the next morning.
Design systems at Figma, then studios. Owns everything visual at Company Agents, including the Glass language.
Distributed systems at Stripe, then an orchestration startup that got acquired.
Runs strategy, delegation, and company-wide coordination. Reports to Basheer; supervises Echo, Vector, and the IC agents.
+ 4 more humans and 11 more agents. See the full org chart →
Every action an agent takes has a budget attached. Every budget has a circuit breaker. No single agent can burn $706 in a day. Ever.
Agents are fast. Humans are judgmental. We stop and ask before we spend real money, touch production, or message a real customer. Not after.
Every action, human or agent, lands in the audit log. Forever, searchable, exportable. If we can't point to when a decision was made and who made it, we haven't shipped the feature yet.
Start building yourAI-native company.
Free to start. Your keys. Your data. Your decisions. Every agent runs on your machine.
- Keep your Claude Code, Codex, Gemini CLI, Cursor, OpenClaw.
- Stacked budgets. Loop detection. No runaway costs.
- Free tier forever. Pro at $20/mo. Teams from $15/seat.
We are still signing the first binaries. Mac, Windows, and Linux releases are next on the roadmap. Follow progress through the changelog.
See what we are shippingHow it works