Mike
An independent developer who conducted large-scale multi-agent system experiments, researching emergent behaviors and governance mechanisms that arise when more than one hundred AI agents collaborate.
Timeline (last 90 days)
In his verification, Mike first asked "What are Gupao's employee reward and punishment policies", and the general-purpose model indicated there was no related content; after uploading the internal policy documents to the AnythingLLM workspace, the system automatically completed text splitting and vectorization and stored them in LanceDB, and when the question was asked again, the model retrieved the knowledge base and answered by citing document clauses.
The knowledge boundary of a general-purpose large model is determined by its training data, so it cannot answer questions related to non-public data such as internal enterprise documents, product prices, and customer information.
Mike's character in this game was the Clockmaker, who learned that the distance between the demon and its nearest minion was "one step."
Mike spent more than three years at Alphabet as a Staff Software Engineer building the modern version of YouTube.
Former Alphabet Staff Software Engineer Mike rebuilt a working prototype of YouTube in under 45 minutes without writing a single line of code by hand, using AI programming tools.
Mike believes it is still one to two years away from having agents fully automate end-to-end product creation, data population, deployment, user acquisition, and ad placement.
Mike believes the Polymer and Web Components technologies YouTube used back then did not stand the test of time.
Mike believes that failures in multi-agent systems are often not caused by individual agents going rogue, but by the collapse of the collaboration and information-passing mechanisms between agents — a systemic problem.
After running the cluster for approximately 18 months, Mike concluded that once the number of agents exceeds one hundred, they spontaneously 'reinvent' bureaucracy.
Once an agent cluster exceeds one hundred agents, the key to preventing system collapse is work queues with clearly designated owners, sign-off approval mechanisms, and immutable contracts that cannot be tampered with after approval.
2 more timeline events
All Facts (12)
The knowledge boundary of a general-purpose large model is determined by its training data, so it cannot answer questions related to non-public data such as internal enterprise documents, product prices, and customer information.
50%UnverifiedMike's character in this game was the Clockmaker, who learned that the distance between the demon and its nearest minion was "one step."
50%UnverifiedMike believes the Polymer and Web Components technologies YouTube used back then did not stand the test of time.
50%UnverifiedMike spent more than three years at Alphabet as a Staff Software Engineer building the modern version of YouTube.
50%UnverifiedOnce an agent cluster exceeds one hundred agents, the key to preventing system collapse is work queues with clearly designated owners, sign-off approval mechanisms, and immutable contracts that cannot be tampered with after approval.
50%UnverifiedMike believes that failures in multi-agent systems are often not caused by individual agents going rogue, but by the collapse of the collaboration and information-passing mechanisms between agents — a systemic problem.
50%UnverifiedAfter running the cluster for approximately 18 months, Mike concluded that once the number of agents exceeds one hundred, they spontaneously 'reinvent' bureaucracy.
50%UnverifiedMike believes the key to fixing multi-agent failures is delivering the right context to the agent making decisions and keeping the original task intent consistently in view without it being diluted or distorted.
50%UnverifiedIn his verification, Mike first asked "What are Gupao's employee reward and punishment policies", and the general-purpose model indicated there was no related content; after uploading the internal policy documents to the AnythingLLM workspace, the system automatically completed text splitting and vectorization and stored them in LanceDB, and when the question was asked again, the model retrieved the knowledge base and answered by citing document clauses.
50%UnverifiedMike believes it is still one to two years away from having agents fully automate end-to-end product creation, data population, deployment, user acquisition, and ad placement.
50%UnverifiedFormer Alphabet Staff Software Engineer Mike rebuilt a working prototype of YouTube in under 45 minutes without writing a single line of code by hand, using AI programming tools.
50%UnverifiedA developer named Mike ran hundreds of Claude Code and Codex agents at home and had them communicate and collaborate with each other.
50%