跪拜 Guibai
← All articles
Algorithm

Qwen3.8-27B Runs 40-Step Agent Tasks Inside DeepSeek Harness

By 致Great ·
Read original on juejin.cn ↗ Google Translate ↗ Alt translation

A 27B open-weight model running locally can now sustain multi-step agentic workflows that previously required much larger proprietary models. The combination of vLLM's tool-call parsing and DSH's execution harness and trace view gives developers a self-hosted, debuggable alternative to closed agent platforms.

Summary

A real-world integration test drops Qwen3.8-27B-FP8 into DeepSeek Harness and assigns it a multi-step WorldQuant alpha research task. The model reads project files, writes and executes Python scripts, launches parallel simulations, and maintains a running task list across nearly 40 steps. When a race condition corrupts the results file, it diagnoses the concurrency bug from logs and recovers by reading the JSONL output instead. The entire session ran continuously for over 30 minutes with no goal drift.

Setup is two-stage: serve the model with vLLM using the official reasoning and tool-call parsers, then register the OpenAI-compatible endpoint as a custom provider inside DSH. The critical detail is that the model ID in DSH must match vLLM's `--served-model-name` exactly. DSH's trace view then exposes every tool invocation, parameter, and timing breakdown, turning an opaque agent run into an inspectable execution chain.

Takeaways
Qwen3.8-27B-FP8 served through vLLM plugs into DeepSeek Harness as a standard OpenAI-compatible endpoint; no custom adapter is needed.
The vLLM launch command requires `--reasoning-parser qwen3`, `--enable-auto-tool-choice`, and `--tool-call-parser qwen3_coder` for tool use to work inside DSH.
The model ID registered in DSH must match vLLM's `--served-model-name` character-for-character, or the harness will report the model as unavailable.
Context length is set to 256K via `--max-model-len=262144`, but can be lowered to 64K or 128K if VRAM is tight at startup.
During a 30-minute, 40-step quantitative finance task, the model recovered from a race condition in result writing by switching to JSONL logs without human intervention.
DSH's trace view breaks the entire agent run into a chronological chain of model outputs, tool calls, context injections, and timing data, making long-running agent sessions debuggable.
Conclusions

Tool-call reliability in open-weight models has reached a threshold where a 27B parameter model can sustain a 40-step agent loop without derailing, which shifts the bottleneck from model capability to harness design and observability.

The Docker GPU device mapping (`--gpus '"device=2"'` with `CUDA_VISIBLE_DEVICES=0`) is a common footgun for self-hosted setups; the container remaps host GPU indices, so the two values intentionally differ.

DSH's trace view addresses the core pain point of agent development: when a model runs unattended for 30 minutes and produces a wrong answer, you can pinpoint exactly which tool call or reasoning step went wrong instead of replaying the entire session.

Concepts & terms
vLLM
A high-throughput inference engine for LLMs that provides an OpenAI-compatible API server, supporting features like continuous batching, PagedAttention, and tool-call parsing.
DeepSeek Harness (DSH)
An agent framework that gives an LLM access to a filesystem, shell commands, and tools, orchestrating multi-step tasks while recording every action in a traceable execution chain.
KV Cache
A memory structure that stores key-value pairs from previous tokens during autoregressive generation. Its size grows with context length, making it the primary VRAM consumer for long-context inference.
Tool-call parser
A vLLM component that extracts structured tool invocation requests from a model's raw output, enabling the model to request actions like running shell commands or reading files.
Source: juejin.cn ↗ Google Translate ↗ Backup ↗