Pion: The future of automation companies or a bubble?

Look at the numbers. A seed-round entrepreneurial team (12 people, ARR $300k) connected to Pion once conducted an A/B test: it handed over the entire process of "releasing new features and launching" to the Agent for two weeks. Result? The PR descriptions generated by agents are three times more detailed than humans, and the test case coverage has increased from 62% to 89%, but the launch delay has increased from 4.2 hours manual to 11.7 hours . Why? Because Agents are stuck frantically cycling tokens on explicit process nodes such as "Generating CHANGELOG","Updating Confluence Documents","Notify Sales Groups" and "Synchronizing Notion Roadmap", the real blocking points-"Waiting for App Store review","Waiting for Database Migration Window" and "Waiting for Legal Confirmation of Compliance Terms"-are all outside the coverage of MCP Tool. Agents don't know how to "wait", they only know how to "do". It explicitly transforms the organization's hidden waiting costs into explicit computing costs, and the final payment is OpenAI's bills and engineers 'overtime pay.

This is the efficiency paradox: The cost of making the process explicit is often higher than the value of the process itself. In the past, when a product manager shouted "Go online," engineers knew which documents to bypass, whom to say hello to, and which tests to skip. Now the Agent strictly abides by the SOP, and every node has to be checked. After the Token is burned out, the task has not yet been closed. The most extreme case I have seen: a three-person team used Pion to run "Weekly Report Generation". Prompt inserted 15 role definitions, and a single call of 120,000 Tokens at a cost of $1.8. No one read the generated weekly report-because everyone still had a weekly meeting to align. This is not an automation company. It is clearly a turbocharger installed on bureaucracy.

At this time, someone always jumped out and said,"Are you looking down on Scaling Law? o1 The reasoning ability of the model is strong, the Context Window is expanded to 1 million, and the MCP ecosystem is mature. Wouldn't these problems be solved?"-- Classic Steelman, but can't catch up with reality.

Strong reasoning ability ≠ judgment. O1 can deduce the 15-step legal process of "how to lay off employees legally", but it does not know that "if you lay off the old engineer who has been writing code for five years now, no one will understand the core architecture, and the next version will fail." This is called tacit knowledge, and it exists in the human brain and team network, not in any documents, code bases, or Slack history. No matter how big the Context Window is, what it can hold is explicit data, but what it cannot hold is "trust" and "tacit understanding." Not to mention the MCP ecosystem: The essence of current MCP Server is to translate API documents into natural language interfaces. It solves "how to tune", not "why to tune" and "who is responsible after tuning". When the Stripe payment callback fails, AWS bills soar abnormally, and the core database switches between master and slave, the Agent can call the Runbook. However, for the step of "Contact the SRE on Duty" written in the Runbook, the Agent can only send Slack @here. The person who really gets up and changes the configuration is still a person.

Put this matter into quantitative trading to see it more clearly. High-frequency trading (HFT) companies have long realized the core aspects of an "automated company": strategy generation, risk control, ordering, and execution without intervention. But why are Citadel and Two Sigma still employing hundreds of PhDs and engineers? Because of policy failures, model drift, regulatory policy changes, and black swan events, no Agent can "reason" these non-stationary distributed decisions from historical data. HFT automates "execution certainty" and retains human judgment of "policy uncertainty". Pion's attempt to automate the latter is like letting an autonomous driving system not only drive, but also decide whether to go to work today and who to sell to when the company goes bankrupt.

I know an architect, Lao Zhang, who spent 7 years in a unicorn. Last year, the company launched a Pion-like system, and the boss said,"Release manpower to innovate." Lao Zhang was originally responsible for core link reconstruction, but what does he do every day now? Modify Prompt to the Agent, write the adaptation layer of the MCP Server, review the low-level code generated by the Agent, and hold a meeting to explain why the Agent's solution "conforms to the specifications but doesn't work." He complained to me: "I used to write code, but now I write code for the AI who writes the code, and I also have to write requirements documents for the AI's product manager. I was not replaced, I was downgraded to an AI operation and maintenance engineer. " This is the real impact of the Pion on culture: is not a substitute for people, but turns senior engineers into low-level prompt engineers, outsources managers 'judgment to probability models, and compresses the organization's fault tolerance rate to the line where "Tokens do not report errors".

To put it bluntly, products like Pion are not sold as "automation companies" but as "a sense of security for anxious founders and investors." Founders are worried that they cannot recruit people, manage people well, and people are too expensive; investors are worried that Portfolio burns money too fast and Headcount is not sexy enough. Buy a Pion, the organization chart is beautifully drawn, the Agent list is neatly arranged, and "AI-Native Organization" is written in the report materials. The next round of financing valuation is easy to negotiate. That's why smart people do stupid things together-- Everyone is optimizing the indicator of "looking like the company", except that "surviving" is the only KPI of the company.

I make a bet: In the next 18 months,"fully automated companies" like Pion will become the AI application category that is "most suitable for doing demos, the most difficult to implement production, and the highest refund rate". What can really work through is to reduce the dimension of the Agent to a "super intern with a single function": help Sales write emails, help Eng write individual tests, help PM compile PRD, and help legal initially screen contracts. As long as it is not allowed to try to "understand the business" and "take responsibility", it is a good tool; once it rises to the abstraction level of "running the company", it is a burning money, creating illusions, and complicating simple problems. ****

One last question: When the Agent can really handle decisions such as "layoffs","M & A negotiations" and "crisis public relations", is it the founder sitting on the board and signing the signature or the engineer who wrote the System Prompt? If even the founder doesn't want to be responsible for this list, why should the model take the blame for you?