AI Tencent Youtu Lab survey: finish tasks, not answers

A survey paper by Tencent's Youtu Lab and several Chinese universities, published Jun 28, 2026, argues that AI must stop producing single-shot answers and start finishing entire tasks inside persistent work environments. The researchers map a shift "from chatbot to digital colleague" that centers on two capabilities: slow, verifiable reasoning and workspace-backed, reusable skills.

How do the authors define the shift from chatbot to coworker?

The paper describes a five-stage evolution that moves models from fast, token-by-token answer production toward delegated task execution that ends in verifiable completion. Early chatbots stored language patterns in parameters and produced answers in one pass; the thinking-LLM era, initiated by OpenAI's o1 and Deepseek-R1, invests compute at inference to explore solution paths, verify intermediate steps, and self-correct, borrowing Daniel Kahneman's System 1 and System 2 distinction for intuition versus deliberation.

These thinking LLMs learn through reinforcement learning to reward verifiably correct solutions, the authors say, shifting the objective from plausible responses to provable correctness and completion.

What are "workspace" and "skill" and how do they enable finishing tasks?

A workspace supplies persistent state across a workflow, while a skill packages operational know-how into reusable bundles; together they let models convert intent into finished work. The paper contrasts fragile first-generation agents that called tools but left no lasting state with what it calls the OpenClaw era, where files, sessions, logs, browsers, permissions, and skills survive across the entire workflow.

The authors point to systems such as OpenHands and SWE-agent that embed agents in controlled development environments and note that Anthropic's Agent Skills formalize skills as folders containing a SKILL.md file with instructions, scripts, and resources. They write that "skills aren't prompts, and they aren't traditional tools either," positioning skills between model reasoning and workspace execution. The survey cautions that skills can become stale, overfit workflows, or create attack vectors, so lifecycle management, sandboxing, permission controls, rollback, and workspace hygiene are required for reliable deployment.

How is evaluation and training different for workspace-based agents?

Training and evaluation move from instruction-response pairs to state-action-observation trajectories, and success is measured by task closure rather than answer accuracy. Benchmarks such as SWE-bench, OSWorld, and WebArena require reproducible starting states, executable tools, trajectory logs, and end-state checks. The paper highlights that GPT-4 initially completed just 14 percent of WebArena tasks, illustrating the gap between static Q&A benchmarks and realistic web environments.

A separate evaluation cited in the paper found skill adoption is uneven: a Vercel test showed coding agents did not call a provided skill system 56 percent of the time, while a compressed documentation index in an AGENTS.md file achieved 100 percent success and the skill system topped out at 79 percent, suggesting passive, always-present context in the workspace can beat active skill retrieval.

Why it matters

If AI is judged by whether it actually finishes work, the engineering problem shifts from bigger base models to the software and operational layers that hold state, manage credentials, and verify execution. Persistent workspaces broaden the attack surface by exposing credentials, local files, identity tokens, and communication channels, turning security into an operational concern. Projects like OpenClaw PRISM and ClawGuard are already proposed runtime safeguards for permissions, provenance, and audit logs, reflecting this new risk profile.

What to watch

Look for benchmarks and papers that report trajectory-based completion rates in realistic, stateful environments, and for platforms that demonstrate skill lifecycle tooling such as versioning, verification loops, and rollback. Concrete progress will show up as meaningful gains in end-state completion metrics on environments like WebArena or SWE-bench and wider adoption of workspace-first patterns in agent deployments.