A coding agent is a long prompt, repeated.

A coding agent resends the system prompt, the repo context, and the conversation on every turn. The slow part is often reading that prefix again, plus a GPU that disappears mid-edit. The token bill is the other part: agents write more tokens than a chat box.

The only code number we will quote

In the 23-way H100 test, Qwen3.8-27B on one H100 did 360 tokens per second on the code prompt and 211 on plain chat. The best code row called out next to it was DekaLLM at 376. One request at a time, one day, thinking off. It is not a measurement of a coding agent, a tool loop, or time to first token for an IDE. We have not published those.

The 50.8 → 213 run is the same checkpoint with FP8 and a draft model, across mixed prompts, still one request at a time. At 16 requests the same H100 did 130 tok/s each and 1,817 tok/s for the card. Speculative decoding helps more on predictable text, which code often is, and less once the batch is full. That post also divides a $4.50-an-hour rental at 213 tok/s into about $5.87 per million output tokens, at one request at a time. That is the rental, not a denv price, and not a price for an agent. denv does not sell tokens. Your GPU bill stays with the provider. See pricing.

What keeps the next turn cheap

Send the next turn to the replica that already holds the prefix. Spraying it across clouds throws the cache away. llm-d does that inside a site. Across sites, the request sticks until the site cannot answer, then fails over. The write-up is one cluster.

A dead GPU is repair: cordon, drain inside the disruption budget, reset, reboot, driver reinstall, or replace a cloud machine. The agent’s next request lands on a healthy replica. A half-written stream is not quietly resumed on another provider.

What is not live

The MCP server, so a coding agent can deploy a model itself, is marked coming soon in the docs. This page does not pretend it is shipping. You can still point an agent at an OpenAI-compatible endpoint you already have.

There is no published “coding agent on H100” recipe with its own latency. Llama, Qwen, and DeepSeek say so when a number is missing. The product is early. Write to anup@denv.ai with the model you actually run.