Today the conversation around agents was less about who launched what and more about how much it costs and who actually uses it. I’ve split the highlight and the rest of the radar below.

This week I read that Garry Tan, who runs Y Combinator, says he spends between 100 and 200 thousand dollars a year on tokens alone to run his agents.

I laughed at the number, but it is the most honest data point in the story. Tan himself admitted that today it is too expensive for a personal agent to become a mass-market thing.

Anyone who works in product recognizes this math. Everyone wants to put an agent inside the product, and almost nobody designs the unit economics first. How much does each completed task cost? Who pays: the user, the plan, the margin? The guide to AI agent cost management and the cost budget template help build that math.

Another detail that caught my attention: his agent sees email, files and notes. The more context, the more useful the agent gets, and also the more cost and the more responsibility for the data. This is where the AI governance guide comes in.

I’m optimistic about this. Cost per token tends to fall, and when it does there will be room for those who have already learned to measure value delivered per task. For the PM, the practical lesson is to start treating inference cost as a product metric, alongside activation and retention. If the topic is bringing agents into your product, also read the guide to AI agents.

For those who got curious and want the full story, here is the full read.

The rest of the radar

  • Conductor is going to Meta — a reminder of the risk of depending on a small startup’s AI tool: the cloud shuts down on 11/20. Conductor
  • Anthropic and FlashProve — cryptographic proofs showing which model produced an answer, useful for audit and compliance; still research. Anthropic
  • Ripple-65B Preview — a US open-weight model for agentic coding, available from several providers; figures are the company’s own. Ricursive
  • Kimi K3.1 Pro with open weights — a multimodal agentic model with 1M context, but with its own license to read before using. Hugging Face
  • DoorDash and outside agents — going by the headline, only hundreds of orders have come from third-party agents, an adoption data point that cools the channel thesis. Ars Technica
  • OpenAI: Images 2.6 and the Pierre acquisition — two OpenAI pages on the HN front page; I couldn’t open the content, so it’s worth checking the official announcement. OpenAI

More tomorrow. If cost is the barrier, the next gain will come from those who measure better.