An AI coding Agent takes 47 seconds to complete a task. How much of that 47 seconds goes to model inference, tool calls, and retry waits? DeepSeek open-sourced its coding Agent runtime Harness (DSH) this week, and Alibaba Cloud's observability team immediately built two monitoring tools for it. We've noticed that the competitive focus of Chinese large-model companies is shifting from "can the model write code" to "how stable, cheap, and traceable it runs in production."
What this is
DeepSeek Harness is DeepSeek's open-source "AI coding Agent operating system" — the model generates code, while Harness manages multi-turn conversations, schedules tools, controls permissions, and dispatches Subagents when needed to collaborate on tasks. It's distributed as an npm package, with the command-line entry point called `dsh`, and Web mode listening on 127.0.0.1:3080 by default.
Alibaba Cloud's LoongSuite is the "dashboard" for this system. Observability is originally an ops concept, focused on correlating three data types — Logs, Traces, and Metrics — to pinpoint where a system is slow or failing. In the Agent context, that translates to: did those 47 seconds go to the model or to tools? How many retries after failure? Is token consumption coming mainly from the main Agent or from subagents?
Alibaba Cloud offers two paths: a standalone installable open-source plugin that sends data over the OpenTelemetry protocol to backends like Jaeger, Tempo, and Langfuse; and another integrated into Pilot, suited for teams running Claude Code, Cursor, DSH, and other Agents side by side who need unified management.
Industry view
Supporters see this as a clear signal that Chinese AI companies have entered an "engineering phase." The willingness to invest in monitoring, billing, and debugging tools means Agents are already running in real production — only when money is being burned and things are breaking do people care where those 47 seconds actually go.
But there are plenty of sober voices. One critique is "over-engineering": most enterprises haven't even gotten their first Agent running, so talking about observability now is like debating rocket recovery before learning to spark a fire. A second concern is "ecosystem lock-in": all monitoring data ultimately lands in Jaeger, Langfuse, or Alibaba Cloud backends — the more Agents are used, the deeper the dependency on this stack, which is just another form of vendor lock-in. The third practical issue: DSH is still at 0.1.0-rc, requires Node.js 22.19.0 or above, and putting it into production means betting on an unstable open-source project.
Impact on regular people
For enterprise IT: Future AI coding tool procurement contracts will include hard metrics like "cost per thousand calls, failure rate, average latency." The era of judging by demo is ending.
For individual careers: Programmers using AI to write code may soon need to watch their Agent dashboard like they watch cloud monitoring — knowing which tasks are quietly burning money and which can be shut off.
For consumer markets: Individual users won't feel any change in the short term, but SaaS vendors' AI coding costs will become more transparent — and over time, that will show up in subscription pricing and feature tiering.