Build Self‑Evolving DeepSeek Agents for Just ¥0.2 with PenguinHarness

PenguinHarness is the world’s first harness that natively supports multiple self‑evolving agents. It lets users build a new agent from scratch for as little as ¥0.2 in token cost, roughly half the time and 1/200 the expense of OpenAI Codex, while delivering results that are closer to production‑ready.

In benchmark tests the tool raised agent accuracy from 50 % to 90 % and, when compared side‑by‑side with Codex on a classic RAG task (generating a chunk‑based retrieval‑augmented answer with citations), PenguinHarness produced smoother, fully streamed output with proper references, whereas Codex’s answer mixed languages and lacked streaming.

The project builds on the same core ideas discussed in the AI‑4‑AI community, such as recursive self‑improvement (RSI). It cites Meta’s former research director Tian Yuandong’s Recursive work, which shows agents can train models and optimise GPU kernels, and MSRA’s former vice‑president Bian Jiang’s XYZ AI Lab, which demonstrated large‑scale agent‑driven deep‑research model training.

PenguinHarness introduces a custom evaluation suite called GDPevo , designed to test self‑evolution across six real‑world domains (medical, finance, legal, etc.). GDPevo separates training and test sets to prevent agents from simply memorising answers during evolution.

The self‑evolution loop consists of four explicit steps:

Organise evaluation – an Optimizer Agent spawns multiple Evaluator Agents to run the benchmark questions.

Independent scoring – each Evaluator runs the target agent, scores it against hidden rubrics, and returns the score.

Analyse and optimise – the Optimizer aggregates results, inspects trajectories, identifies loss causes, and creates a candidate next‑version agent by adjusting prompts, skills, or configuration.

Validate iteration – Evaluators re‑score the candidate; the Optimizer accepts it only if the score strictly improves, otherwise it rolls back.

This process is visualised in the article’s diagram and ensures that only genuinely better versions are retained.

To keep evolution safe, PenguinHarness enforces a “contract” that limits modifications to prompts and skills, requires snapshot/rollback capability, hides rubrics from the target agent, and records all optimisation steps as files for auditability. The contract is formalised in a CONTRACT.md file.

Beyond the core loop, the harness bundles many useful features: built‑in skills for LlamaFactory, vLLM, Ollama, a unified model gateway supporting over 1,000 online and local models (including Kimi K3, Gemini 3.6, Ling 3.0 Flash), and a visual‑modality extension that lets DeepSeek see generated webpages or PPTs and adjust based on visual feedback.

It also provides fine‑grained trajectory analysis, a minimal shell‑based agent template comparable to Pi Agent, and a dramatically smaller system prompt (≈1,300 tokens vs. Claude Code’s 15,000).

In production, two deployments are highlighted: a medical‑report‑verification agent that reduced review time from 30 minutes to seconds, and a manufacturing‑line inspection agent that cut downtime by 65 % and nearly doubled output.

PenguinHarness is released under the Apache 2.0 license, runs on Linux, macOS, and Windows, supports one‑click installation, multi‑user isolation, and can be accessed via browser or deployed locally as a low‑cost alternative to Codex.