HKUDS details OpenSpace, a skill self-evolution engine for AI agents

HKUDS has published a detailed walkthrough of OpenSpace, an open-source skill engine that lets AI agents stop starting every task from scratch and instead accumulate working patterns and reuse them. The material shows the full cycle: from configuring an OpenAI model to comparing a "cold" and a "warm" run, where reusing skills reduces token consumption and improves answer quality.

From a cold start

OpenSpace is installed from a GitHub repository, and the tutorial configures it with OpenAI models, including gpt-4o-mini. The agent is then given a first task with no ready-made skill library. The system executes it like an ordinary LLM agent, but in parallel records successful steps, workarounds and working instructions that can be reused later.

Skills do not live only in the conversation context: they are saved as files following the SKILL.md convention and written to a SQLite database, so they can be browsed, versioned and analyzed separately from the run itself.

The authors then run a similar task again and show the difference between cold start and warm start. Here OpenSpace no longer reasons from scratch but selects suitable skills through hybrid search using BM25 and embeddings. The demo also creates three baseline skills by hand: data validation, report generation with fallback scenarios, and an error-recovery mechanism. The idea is to seed useful patterns in advance and let the evolution engine grow them on real tasks.

How skills grow

The key idea of OpenSpace is that a skill is treated not as a static prompt but as a living entity. After each execution, the system analyzes what worked, where failures occurred, which tools degraded and what could be improved in the next version. So the engine not only accumulates successful scenarios but can also repair them, specialize them for new tasks and extract new skills from single successful runs.

• FIX — repairs a broken or outdated skill without changing its role.
• DERIVED — creates a derived version for a narrower scenario or a new class of tasks.
• CAPTURED — extracts a new reusable pattern directly from a successful execution.
• BM25 + embeddings — helps quickly find the most relevant skill for a task description.
• open-space.cloud — provides a shared catalog where skills can be searched, downloaded, uploaded and shared between teams.

An important part of the design is the collective layer. Through the cloud community, agents can exchange skills that have already evolved, view version history and build shared repositories for a team. The article presents this as a shift from a single assistant to a network of agents that learn not only from their own mistakes. If one agent finds a reliable workaround for PDF generation, table parsing or web research, another can take that skill and avoid repeating the same cycle of trial and failure.

OpenSpace token economics

The strongest part of the material is not the setup but the numbers. The OpenSpace repository cites the GDPVal benchmark on a sample of 50 professional tasks in six categories: documents, compliance forms, media, engineering, spreadsheets and strategy. For a clean comparison, OpenSpace was measured against the baseline agent ClawWork on the same backbone model, Qwen 3.5-Plus, so the difference is explained by skill accumulation rather than a model swap.

The result: 4.2x higher economic return, 70.8% average quality and 45.9% fewer tokens when tasks are re-run.

The per-category breakdown shows where self-evolution pays off most. In documents and correspondence, the warm run cut token consumption by 56%; in forms and compliance tasks by 51%; in media by 46%; and in spreadsheets by 37%. The authors also stress that of the 165 evolved skills, most relate not to narrow domain expertise but to execution reliability: handling file formats, error recovery, document generation and quality checks. In other words, the main gain comes not from "industry knowledge" but from the agent no longer breaking in typical technical spots.

What this means

OpenSpace shows well where agent frameworks are heading: from one-off prompts to persistent working memory, where every completed task makes the system cheaper and more robust. For product teams, it is a signal that the next wave of efficiency will come not only from new models but also from the infrastructure of reusable skills around them.