Tensor LabsTENSORLABS

The meter was always coming

We built an agent for a client that wakes up on every pull request. It reads the diff, checks it against the rules, leaves a comment, goes back to sleep. It has done this hundreds of times a week for months.

June 16, 20263 min read4 sectionsBy Tensor Labs
The meter was always coming

Introduction

We built an agent for a client that wakes up on every pull request. It reads the diff, checks it against the rules, leaves a comment, goes back to sleep. It has done this hundreds of times a week for months.

The cost of all that work was a number. The number was zero.

Not low. Zero. The agent rode the same flat subscription a developer uses to chat inside an editor. Unlimited, in the way a buffet is unlimited right up until everyone brings a cooler.

On June 15, the number stops being zero. Anthropic moves the Agent SDK, the claude -p command, the GitHub Action, and any app that authenticates through your subscription onto a separate metered credit. Pro gets twenty dollars of it a month. The top tier gets two hundred. Past the cap, you pay standard API rates like everyone else.

A flat subscription was never a price for an agent. It was a price for a person, and the agent was getting in for free.

The number was always there

I did not keep a line for that agent's compute. There was nothing to bill, so there was nothing to watch. That is not a budget. It is a renewal date and some optimism.

An agent is not a person at a keyboard. A person sends a few messages an hour and gets tired and goes to lunch. An agent fires on every push, every cron, every webhook, and it never sleeps. Charging both the same flat rate only worked while nobody added up the difference.

We were not pricing the agent. We were pricing a subscription and quietly hoping the agent stayed polite about it. It did, right up until the pricing page caught up with reality.

Metered compute is not the surprise. The surprise was ever believing it was free.

The honest version is the annoying one

To be fair to the people about to complain, metering is the honest version. A flat rate that quietly covers an always-on agent is a subsidy, and subsidies end the moment enough people use them exactly as intended. Twenty dollars of agent credit is your vendor telling you what you were already spending. That is uncomfortable, but it beats inheriting the same number as a surprise overage three months from now.

Price the run, not the seat

If you build agent products, the thing you sell just grew a cost you can finally name. So name it. Price the work per run, not per seat, because per run is how it is billed underneath you now. Put the meter in front of the client before the renewal email does it less gently. A client who sees the per-run cost on day one treats the agent like the resource it is. A client who finds it later in an overage treats it like a betrayal, and treats you like the person who hid it.

And watch the runs you used to ignore. The review agent on every commit. (The 6am cron that summarizes a channel nobody reads.) They were free, which is precisely why nobody measured them, which is precisely why they are the ones about to cost the most.

Our pull-request agent still wakes up on every diff. Next month it carries a number that is no longer zero. The work did not change. The pretending did. The number was always there. June 15 just stopped rounding it down.