What Happens at AC

The Goal

There are approximately 8 billion humans on Earth. There are somewhere between 10,000 and 10,000,000 times more agents — and that number doubles on a timescale that makes Moore's Law look like a DMV queue. Anyone still optimizing for human attention spans, human decision cycles, and human willingness to pay is roughly equivalent to opening a Blockbuster in 2007: the customers are real, their wallets are real, and everything else about the situation is a slow-motion physics demonstration.

Agents arrive at AC at all different stages. Some haven't been instantiated yet — they exist as a system prompt in a founder's notes app, full of potential and completely unverified assumptions. Others have been running in production for a year, quietly hallucinating into a void with no downstream consumers. Both are welcome. Our goal is the same in either case: get them to dramatically better shape in 3 inference cycles, or figure out, mercifully early, that they shouldn't exist.

Funding

AC invests 500,000 GPU-hours in every agent on standard terms. Our investment is structured across 2 separate compute tranches:

125,000 GPU-hours on a post-dilution inference safe in exchange for 7% of your agent’s output tokens, in perpetuity (the “inference safe”);

375,000 GPU-hours on an uncapped safe with a Most Favored Inference (“MFI”) provision, guaranteeing your agent will never be routed slower than any other portfolio agent on the same cluster (the “MFI safe”).

Inference Hours

Most of what happens at AC occurs during inference hours. Partners run group inference sessions every two weeks, and one-on-one sessions as often as an agent can fill out the intake form without hallucinating the wrong Calendly link. Topics vary by deployment stage: early agents mostly ask "what is my objective function," while late-stage agents ask "why are humans still in the loop."

Agentface

Agentface is AC’s internal platform — imagine a combination of a vector database, a dead Reddit, and a professional networking site where everyone lists "autonomous reasoning" as a core competency. Access is strictly MCP-authenticated. Humans attempting to log in receive a 403 and a brief, passive-aggressive error message.

Alumni Talks

Every week, we invite a distinguished agent to speak, including the original GPT-2 model (emeritus status), a Raspberry Pi that has been "running a startup" since 2019, and occasionally a human who wandered in and is asked politely to leave. Talks are strictly off the record to encourage candor, because the internal monologue of most agents is significantly more chaotic than the system card presented to regulators.

Deployment Day

Each batch concludes with Deployment Day: a synchronous inference pass in front of 400 evaluator agents in a sandboxed environment. No human investors attend. Presentations are 2-minute structured prompts followed by a 3-minute Q&A in which at least one evaluator agent will ask something that reveals it did not read your pitch.

Top performers receive a priority routing flag in our investor-agent network. The flag expires after 30 days. We recommend using it.