1Context and scope
A coding agent takes a task and loops: think with the model, call a tool (read a file, grep, run tests in a sandbox), look at the result, repeat. Each turn is a model call plus a tool call, so a task multiplies into many calls, and any failed step can fail the task. The harness around the loop checks every tool call against a policy, compacts the context when it grows, checkpoints progress so a crash can resume, and traces every step. This design covers one such platform at browser scale, based on Anthropic’s engineering posts on building effective agents and long-running harnesses.
2Goals and non-goals
Goals
- 2 tasks a second, each about 8 turns: a model call, a policy check, a tool call, a sandbox run on 40% of turns, a compaction on 15%, a checkpoint and a trace per turn.
- Success at least 95% of tasks, p99 at most 180 s.
- A model outage does not fail tasks: calls fall back to a second provider.
Non-goals
- The editor and terminal integration.
- Repository hosting and CI.
- Permission prompts shown to a person.
3Capacity estimate
| Quantity | Value | How |
|---|---|---|
| Model calls | ≈ 16/s | 2 tasks × 8 turns |
| Streams held | ≈ 16 × 5 s ≈ 80 | first token 0.6 s plus 400 tokens at 90 a second; 600 streams available |
| Tokens a minute | ≈ 11.5M of a 200M quota | 16 × 12,000 × 60; 9,000 of each 12,000 are cached prefix (not counted) |
| Sandbox runs | ≈ 6.4/s at 1.5 s | 40% of turns: about 10 sandboxes busy of 400 |
| Compounding errors | 1 − (1 − 0.002)^8 ≈ 1.6% | tool errors over 8 turns; the goal allows 5% |
| Checkpoints and traces | ≈ 16/s each | one write per turn |
The rest of the doc
The design · Key concepts · Alternatives considered · Failure modes and fixes · Try it in the lab · Check yourself · Further reading
Unlock the Coding agent platform design doc
Every component and why it is there, the capacity numbers behind it, what breaks when you push it and how to fix it, and the same system open in the lab to break and fix yourself. Paid once, on your account.