Design the backend of a Claude Code-style coding agent from an empty canvas: agent turns that call a model, tools and sandboxes step by step, checkpointed, at a sane token bill.
Unlock it once for 25 credits; everything inside it is free (runs, changes, Submit, results, hints). Sign in with Google to start; new accounts get free credits.
What it must do
Run an agent turn: the model plans, calls tools (read, edit, grep), runs tests in a sandbox, and repeats for a few steps.
Save a checkpoint after every step so a session resumes after a crash or on another device.
Keep the repository context cached at the model so long prompts are not paid for in full every step.
Keep working when the model provider slows down or rate limits.
Scale
8 requests a second: 7.2 agent turns and 0.8 session resumes (about 620,000 turns a day).
Each turn makes 3 model calls: about 22 model calls a second with 12,000 input tokens each (10,000 of them the cached repository prefix) and 80 output tokens.
About 30% of steps run code in a sandbox for around a second.
Request types to route: turn (90%, write), resume (10%, read). Budget: $900,000 a month.
Estimate first
Model calls a day (million)
Input tokens a day (billion)
Monthly input-token bill without prefix caching ($3/M) (USD)
Monthly cloud cost of your design (USD)
The test cases
You start from an empty canvas. When you run the tests, each one simulates your design with its own traffic and failures:
Test
What it does
Passes when
Normal day
The normal day: the traffic in the brief, nothing broken.
Success ≥ 98%; p99 latency ≤ 8.0 s (turn); p99 latency ≤ 500 ms (resume)
Launch day
At 5 s a launch triples traffic for 8 s, past the main provider’s rate limit.
Success ≥ 95%
The model provider slows down
At 5 s the main model takes 8× longer to start answering, for 8 s (an overloaded region).
Success ≥ 97%; p99 latency ≤ 10 s (turn)
Budget: $900,000/month
The normal day's design must cost at most $900,000 a month.
Cost ≤ $900,000/month
Must include
An agent loop. Something runs the steps: ask the model, run the tools it picks, feed results back.
Tools and a sandbox. The agent reads and edits files and runs tests, isolated from everything else.
A checkpoint store. Long sessions must survive crashes and resume where they stopped.
A model. Every step is a model call with a long, mostly repeated context.