CREAO: A Company of About 25, Coordinated by Agents

Key points

• Architects choose which tasks to pursue from the ones AI proposes after reading the code and logs.
• Every pull request goes through three AI reviews, then a person checks for strategic risk.
• With no QA team, three judge models score live output, and people spot-check their verdicts.

Building a system that coordinates, instead of people who coordinate

Software startups have a layer that connects different people. A product planner writes the requirements and tracks progress. An engineering manager hands out tasks to developers and runs regular check-ins. When the code is done, quality assurance (QA) testers, whose job is to find defects and verify software before release, check the features in a separate test environment. Even a slightly larger organization soon fills its calendar with meetings and status reports.

CREAO builds a cloud agent platform, and it decided not to fill this layer with people. Its main product, Super Agent, builds workflows that run on a schedule, keep the memory of earlier runs, and handle many kinds of work across marketing, coding, and video production. According to co-founder and Chief Product Officer (CPO) Clark Gao, CREAO’s business focus started with ordinary consumers and has now moved toward small and medium businesses. Large enterprises are a later target. Plans are credit-based, from a free tier to a $1,000-a-month Ultra tier for teams of up to 50.

On the outside, CREAO targets work that small companies now give to assistants, marketers, outside agencies, and automation tools. On its website, the company compares itself with tools such as Zapier, n8n, Lovable, and Manus.

But the more interesting part is how the company runs inside. CREAO did not start out this way. It first sold a tool for building applications from natural-language prompts, then switched to an agent platform. In February 2026, co-founder and Chief Technology Officer (CTO) Peter Pang overhauled both the product architecture and the way the team works. The roles that changed most were project managers, engineering managers, QA testers, and the meetings that connected them. By the company’s account, agents now handle most of that coordination.

How work flows in an organization of about 25

CREAO describes itself as “Around 25 people, mostly recent grads, no middle management.” In an April 2026 post, Peter Pang wrote: “Twenty-five employees, 10 engineers.” According to the company’s own interview summary, “agents handle the coordination, the status tracking, scheduling, and hand-offs.” Clark Gao describes the sales and customer-facing side as “close to a consulting or law firm partnership model.” Each channel owner brings in work, builds a small team, and owns the result, and leadership collects the results: “We aggregate what each of them delivers.” Gao says new hires sometimes tell him, “I don’t even know who my manager is,” and still deliver.

Peter Pang splits the company’s roles into two kinds. The first is architects, only one or two people in all. Architects build the system design, the test environment, and the standard procedures agents must follow: “They design the standard operating procedures that teach AI how to work.” The second is everyone else, the operators. Operators do the hands-on work. They investigate the causes of bugs, refine the user interface (UI), review pull requests, and verify fixes. Pang puts the direction of work bluntly: “AI assigns tasks to humans.” In Gao’s words, the layer that shrinks is “the roadmap-and-status-update layer.”

The change began on February 5, 2026. Sylvain Laporte, who calls himself Chief AI Wrangler, noticed that code written by a new model needed no touch-ups from people, and he told Peter Pang. The rebuild started that day. CREAO merged about 50 scattered repositories into one huge repository, a monorepo, so that agents could see the whole system at once. It also wrote the team’s knowledge into “skills,” short playbooks an agent reads before it builds a feature. Laporte says, “the codebase should be something you teach.” Arky Yang, an engineer who joined recently, says “everything for engineers is in the codebase” and adds, “I onboarded myself without too much help.” Instead of grabbing a colleague to ask questions, he learned the code by asking a coding agent.

According to Founder Park, which interviewed the founders in April 2026, the engineering loop has six steps.

The first is plan. AI reads the codebase, the production logs, and the harness data. Then, in about 30 minutes, it drafts a plan and proposes the work that is needed.

The second is decide. An architect picks which proposals to pursue and whether they fit the product’s direction. As Pang puts it, “architects decide which to advance.”

The third is build. Coding agents produce the work in a few hours. Pang says AI writes 99% of the company’s production code.

The fourth is review. Every pull request goes through three AI reviews in parallel, checking code quality, security, and dependencies between libraries. After that a person steps in, and the person checks for “strategic risk, not line-by-line correctness.”

The fifth is release. A six-step deployment pipeline builds, deploys, and tests in a dev environment, then moves the code to production. “No phase is optional. No manual overrides.” The company says it ran three to eight production deploys a day for 14 days.

The sixth is learn. Real production logs and the results of A/B tests, which show two versions to users and compare their reactions, feed back into the next plan.

CREAO's six-step engineering loop

One thing stands out. There is no QA team, and no staging environment, meaning a final check environment before deploying to the live service. Instead, a safety system that Pang calls the Agent Harness is running. Three judge models, from Anthropic, OpenAI, and Google, score the replies that agents give users in the live service. Every morning, an agent groups the low-scoring replies and production errors, ranks the groups by severity, and creates work tickets. For the most severe groups, an agent traces back through the code and logs to find the cause, attaches a bundle of evidence, and passes the ticket to a person. For urgent, high-confidence cases, an agent opens up to three draft fixes per run. Changes that touch secrets, CI settings, or access-permission policies are designed to be closed automatically. Before closing a ticket, the system checks six hours of production logs, and if the problem comes back, it reverts the fix. Major changes first go out to only about 10% of total traffic, and the rollout widens only if the judge models’ scores hold. People sample the judge models’ verdicts. If a lasting gap appears between the judges and human judgment, they treat it as a bug in the scoring rubric and fix the rubric.

Areas outside development work the same way. According to Pang, product release notes are “AI-generated from changelogs and feature descriptions,” and social media posts, feature videos, and analytics reports are also made or orchestrated by AI. Candice, CREAO’s social media manager, is in her first job. In about a month, she built four chains of agents. One reviews recent posts every day. One monitors about 20 competitor accounts every week. One remembers the brand voice and drafts replies. The last is a content agent that reads what was trending over the past 24 hours, suggests angles, and makes drafts and images. Candice chooses the angles she likes and approves the final publication. “I’m there to agree or disagree with what they recommend.” Clark Gao says the database is connected directly, so “anyone can just ask the agent about our daily operations,” instead of setting up a meeting with a product manager or a data engineer.

Not every attempt worked at first. Peter Pang told Unite.AI that one content pipeline ran automatically for two days, but the output was poor, so the team shut it down.

Five points where people make the decisions

The biggest difference between a typical software company and CREAO is where coordination and review happen. In an ordinary company, people carry the coordination. A planner writes the spec and tracks the schedule. A development manager divides up the work and runs weekly meetings. But by CREAO’s own account, agents handle status tracking, scheduling, handing off work, proposing tasks, the first code review, and most of the testing.

Instead, people stand only at five clearly named decision points. First, an architect approves what to build. Second, a reviewer checks each pull request for strategic risk. Third, an engineer confirms the diagnosis before an agent makes a fix. Fourth, a person spot-checks the judge models’ verdicts against their own. Fifth, a channel owner approves what gets published outside the company. Peter Pang says that thanks to this structure, the time he spent managing people fell from 60% to below 10%.

Authority sits in a different place too. With the middle layer gone, experienced senior people write the procedures and test environments agents follow, instead of reading status reports. Hands-on staff are fully responsible for the areas they own. Clark Gao says “a sharp junior person with the right tools can own a whole channel,” and Peter Pang wrote that “Junior engineers adapted faster than senior engineers.” The lines between jobs also fade. Gao, who trained as a data scientist, says he did “all the design work for our conference materials” himself, and an intern was told on arrival: “whatever part of CREAO you think you can contribute to, go do it.”

The release rhythm changed a lot as well. Because there is no QA team or staging environment, CREAO deploys the product several times a day and leaves the final evaluation to live production scores and A/B tests. According to Founder Park’s translation of the interview, Peter Pang said: “Ship everything, let AI implement, and let AB test results decide.” This way of working has a clear price. Some product defects are not filtered out by people before release. They are caught afterward by a monitoring system, the harness, on a small slice of production traffic.

Questions for older industries, and four lessons

If this approach holds, the middle layer of small software companies could get thinner. If agents can track progress, assign tasks, and finish the first review, the need for people who only coordinate shrinks. CREAO’s founders say this explicitly. But whether the structure will last once headcount grows past 25 is still unknown. Clark Gao himself adds that “The guidelines and best practices will come as we go.”

Firms that handled marketing and operations for small businesses could also come under pressure. The product CREAO sells is the same as the tool it uses inside. That is, one person strings together chains of agents to do the work that a small team or an outside vendor used to do. Peter Pang says one agent replaced what used to be a three-person search engine optimization (SEO) workflow. If small companies start using similar tools, some of the marketing and operations work that went to outside vendors could be absorbed inside the company.

What other organizations can learn from CREAO is not which AI tools it picked, but how it structured the work.

First, do not bolt AI onto the old flow. Redesign the whole loop of work. Peter Pang says of most companies, “They added AI to the loop. They didn’t redesign the loop.”

Second, write procedures where agents can read them, not only people. CREAO’s monorepo and skill documents give agents and new hires a base to learn the whole system on their own, without a manager explaining it each time.

Third, give clear names to the points where people must decide. Plan approval, risk checks, fix confirmation, and judge-model calibration each have a named person responsible. Putting people at these specific decision points is what fills the gap left by the management layer.

Fourth, count on the bottleneck moving. Sylvain Laporte warns that writing code faster is useless if testing and release cannot keep up: “You moved the bottleneck downstream.” This matches Peter Pang’s point that once building is fast, “a weeks-long planning cycle becomes the constraint.”

What is still not fully proven

A few limits should be kept in mind with this case. Most of what is known about CREAO comes from founders who have to promote the product and from the company’s own blog posts. The figures that AI wrote 99% of the code, or that the company deploys three to eight times a day, were released by the company itself.

Even basic facts differ between sources. The headcount is 20 in the April 2026 interview and 25 in other posts. The founding date also varies by source, from 2023 demos to early 2025. The agent-centered way of working began in February 2026.

There is also no outside study that objectively measures how shipping to real users without a separate QA step affects software quality, error rates, or user experience over the long run. This operating model started only recently and is staffed mostly by young people who recently graduated. As the company itself admits, the detailed rules of how they work are still being written.

---

Sources

1. CREAO — Behind the Agents: What Running an AI-Native Company Looks Like, with Co-Founder & CPO Clark Gao — https://creao.ai/blog/behind-the-agents-ai-native-company — 2026-07-15
2. Peter Pang (X) — Why Your “AI-First” Strategy Is Probably Wrong — https://x.com/intuitiveml/status/2043545596699750791 — 2026-04-13
3. CREAO — The Self-Healing Agent Harness — https://creao.ai/blog/the-self-healing-agent-harness — 2026-05-26
4. CREAO — Behind the Agents: The Cost of Shifting to AI-Native, with Chief AI Wrangler Sylvain Laporte — https://creao.ai/blog/cost-shifting-ai-native — 2026-08-03
5. CREAO — Behind the Agents: How Candice Uses AI for Content Creation to Run CREAO’s X Channel — https://creao.ai/blog/behind-the-agents-how-candice-uses-ai-for-content-creation-to-run-creao-s-x-channel — 2026-08-13
6. CREAO — Behind the Agents: How AI Is Changing Software Development, with Engineer Arky Yang — https://creao.ai/blog/behind-the-agents-how-ai-is-changing-software-development-with-engineer-arky-yang — 2026-10-20
7. CREAO — Behind the Agents: How to Make Demo Videos with AI in Minutes, with AI PM Intern Daksh Kumar — https://creao.ai/blog/behind-the-agents-how-to-make-demo-videos-with-ai — 2026-07-22
8. CREAO — Pricing page — https://creao.ai/pricing — accessed 2026-10-22
9. Unite.AI — Peter Pang, Co-Founder and CTO of CREAO – Interview Series — https://www.unite.ai/peter-pang-co-founder-and-cto-of-creao-interview-series/ — 2026-05-29
10. Founder Park — CREAO: 20-Person Team Ships 8 Features Daily After Pivoting the Organization — https://fpfounderpark.substack.com/p/creao-20-person-team-ships-8-features — 2026-08-19