Subscriptions Priced for Human Pace Meet Agents That Don't Work at Human Pace

On August 11, Capy co-founder Justin Sun announced V2 on X. Capy is a six-person coding-agent company in San Francisco, from Y Combinator's fall 2024 batch. Its homepage describes the product as "Orchestrate a fleet of cloud coding agents in all of the apps your team already lives in." Capy says 70,000 engineers use it. That is the company's own number, with no third-party data behind it. Capy now connects to Slack, GitHub, and Linear. Agents can start automatically from events, split a request across several sub-agents that work in parallel, and hand the resulting pull request to a review agent. Each agent has its own cloud machine. On Capy's pricing page, the Pro and Max tiers say "Use ChatGPT, Copilot, and SuperGrok subscriptions." Users can plug in subscriptions they already pay for and run these agents on that quota. Those subscriptions are billed per person per month, and the price corresponds to one person's usage over a month. When a Capy agent starts working, and how many run at once, depends largely on what happens in the codebase. How today's tools are used With the generation of coding tools that includes Claude Code, Codex, and Cursor, the most common pattern is to open a session, have a person state the request, let the tool change the code, and have the person check it. These tools are also moving to the cloud and to asynchronous work, so a task can keep running after the person leaves. But a person usually starts the task, and the tool mostly owns the step of writing code. Before a problem gets solved, someone has to report it, perhaps as a Linear issue, a Sentry alert, or a line in Slack. Someone then decides what the problem is, who should handle it, and how urgent it is. Afterward someone opens the pull request, reviews it, runs the tests, and merges. Every handoff waits for the next person. As writing code gets faster, waiting takes up a larger share of the time from report to merge. Speed is hard to measure How much AI coding tools improve productivity is still mostly answered by users' own impressions. In the randomized controlled trial METR published in July 2025, 16 experienced open-source developers worked on 246 real tasks. The group allowed to use AI was 19% slower, and afterward they estimated they had been 20% faster. The study is small, and its conclusion may not hold everywhere. I discussed it in my piece on engineering metrics (Chinese). GitHub and Linear record when an issue is created and picked up, when a pull request is opened, and when the first review lands, so the wait in each segment can be computed directly. Capy's main features act on those segments. Tasks that start from events