Source-linked AI summary
From Chatbot to Digital Colleague: The Paradigm Shift Toward Persistent Autonomous AI
Yongheng Zhang, Ziang Liu, Jiaxuan Zhu, Shuai Wang, Xiangqi Chen, Haojing Huang, Jiayi Kuang, Siyu Chen, Ao Shen, Hao Wu, Qiufeng Wang, Qian-Wen Zhang, Junnan Dong, Wenhao Jiang, Ying Shen, Hai-Tao Zheng, Yinghui Li, Di Yin, Xing Sun, Philip S. Yu
TL;DR
LLMs are moving beyond conversational answers, but dependable persistent work still requires stronger cognition and execution infrastructure. This survey frames that shift as a transition to Digital Colleagues built from Thinking LLMs, persistent workspaces, reusable skills, auditable evaluation, and governance, concluding that reliable autonomy depends on these integrated components.
Problem
The paper addresses how LLMs can progress from conversational systems toward dependable agents that reason, act, remember, and complete open-ended tasks.
Method
The survey organizes this transition across Thinking LLM cognition, OpenClaw execution, Workspace + Skill abstractions, data construction, evaluation, and governance.
Results
The survey concludes that self-evolving AI ecosystems can convert operational experience into reusable skills, updated memories, stronger verification, safer policies, and more reliable work outcomes.
Takeaways & Limitations
Reliable digital colleagues require persistent environments, reusable skills, auditable evaluation, and safety governance rather than episodic responses and ad hoc tool calls.
Takeaways & Limitations
Workspace + Skill is a useful lens rather than a complete solution to reliable autonomy, increasing the need for lifecycle management, security review, and operational discipline.
Abstract
from arXiv · showhide
Large Language Models (LLMs) are undergoing a fundamental transformation from conversational generators into integrated AI systems capable of reasoning, action, memory, and self-improvement. We conceptualize this transition as a shift from Chatbot to Digital Colleague: from conversational answers to persistent work. We organize this transition along two tightly coupled dimensions. First, at the cognitive core level, LLMs are advancing from Chatbot-era "fast thinking" systems driven by next-token prediction toward Thinking LLMs that leverage inference-time computation, Chain-of-Thought reasoning, reflection, process supervision, and reinforcement learning to support more deliberate and reliable cognition. Second, at the tool-augmented task execution level, LLMs are progressing from tool-calling Agents that invoke external resources in an ad hoc manner toward OpenClaw-style workstation systems (OpenClaw) equipped with persistent Workspaces, skills, verification loops, and governance. The "Workspace + Skill" paradigm makes episodic tool use colleague-like via state persistence, reusable procedures, task closure, and experience reuse. We examine data construction shifts from instruction-response pairs to State-Action-Observation trajectories and evaluation from static benchmarks to sandboxed, auditable, self-evolving AI ecosystems.
8 Conclusion
The paper concludes by framing next-generation LLM development as an evolutionary roadmap from conversational chatbots toward reasoning cores, tool-using agents, and persistent workspace systems. The timeline organizes this progression by release month and distinguishes open-source/open-platform systems from closed/commercial systems.
- The roadmap depicts an evolution from simple conversational chatbots to reasoning cores, tool-using agents, and persistent workspace systems.
- Each roadmap node is labeled by its release month, situating system capabilities within a chronological evolutionary timeline.
- The figure differentiates open-source or open-platform systems from closed or commercial systems.
1 Introduction
The survey frames persistent autonomous AI as a two-dimensional shift from Chatbot to Digital Colleague: cognitive cores advance toward Thinking LLMs, while task execution advances from Agents toward OpenClaw systems. Workspace + Skill enables durable work, motivating new trajectory-based data, task-closure evaluation, and governance for self-evolving ecosystems.
- Two-dimensional framework: The framework couples cognitive-core evolution from fast Chatbot-era generators to Thinking LLMs with tool-augmented execution from Agents to OpenClaw systems.The cognitive dimension emphasizes reasoning, verification, and consistency; the execution dimension emphasizes action in dynamic external environments.
- Workspace + Skill: Workspace + Skill turns chatbot-style interaction into durable digital-colleague work through persistent environments and reusable procedures.Workspaces retain operational resources, while skills encode planning, tool sequencing, checks, recovery, and validation.
- Data and evaluation: LLM development shifts from instruction-response pairs toward state–action–observation trajectories and task-closure evaluation as systems perform persistent work.Reasoning models additionally use Chain-of-Thought traces, process supervision, and verifiable rewards.
- Structural bottlenecks: Persistent autonomous systems remain constrained by ungrounded reasoning, brittle long-horizon toolchains, transient memory, and safety risks from executable actions.These bottlenecks make reliability, memory, safety, governance, skills, and system maintenance central concerns.
- Socio-technical roadmap: The survey’s roadmap treats Digital Colleague systems as socio-technical systems whose implications include human–AI collaboration, ethics, privacy, creativity, and asset boundaries.The survey also identifies governance and system maintenance as part of reliable autonomous AI.
- Self-evolving systems: Next-generation generative AI is characterized as self-evolving ecosystems that convert operational experience into reusable skills, updated memories, stronger verification, safer policies, and more reliable outcomes.These ecosystems integrate models, workspaces, tools, skills, memories, evaluators, and governance mechanisms.
2 Part I: The Evolution of LLM’s Cognitive Core
The cognitive core of LLMs is shifting from fast, stateless, single-pass response generation toward Thinking LLMs that use extended inference, exploration, reflection, and reinforcement learning for deliberate reasoning. This transition addresses the limitations of pattern matching and enables increasingly agentic behavior through tool use and environmental interaction.
- Chatbot Era: Chatbot-era LLMs served as fast-response interfaces, mapping natural-language prompts to fluent answers through compressed parametric knowledge and single-pass autoregressive generation.This paradigm combined large-scale language generation, implicit knowledge compression, behavioral alignment, and multimodal expansion.
- Limits of Fast-Response Cognition: Fast-response cognition relies on surface-level pattern matching and struggles with rigorous multi-step deduction, causing local errors in mathematical proofs and long-horizon code generation.Autoregressive decoding lacks the lookahead, backtracking, and global search needed for hard mathematics, coding, and planning.
- Thinking LLM Era: Long Chain-of-Thought extends reasoning through coherent multi-step derivations, branching exploration, and reflection within a single generation.These capabilities expand Transformer expressiveness and internalize search previously delegated to structures such as Tree-of-Thought and Graph-of-Thought.
- Thinking LLM Era: Thinking LLMs scale inference-time computation by generating longer reasoning chains or sampling parallel paths, allowing additional test-time computation to improve performance.The passage reports that, with sufficient inference-time compute, a 1B model can surpass a 405B model on mathematical benchmarks.
- Thinking LLM Era: Reinforcement learning in OpenAI o1 and DeepSeek R1 internalized deep, branching, self-correcting reasoning, while unified systems made reasoning an adjustable mode alongside tool use and environmental interaction.This fusion gives modern Thinking LLMs rudimentary agentic behavior and raises the need for architectural support beyond a powerful brain.
- Thinking LLM Era: Evidence suggests reasoning capabilities are latent in pretrained models and may require activation rather than creation from scratch.This hypothesis has implications for the efficiency of reasoning training.
3 Part II: The Evolution of Tool-Augmented Task Execution
Tool-augmented task execution has evolved from isolated tool calls toward grounded, trajectory-level policies and standardized infrastructures, exposing the Agent era’s limits. OpenClaw-style systems address these limits through persistent workspaces, reusable skills, task closure, verification, and governance, while introducing new reliability and security risks.
- Agent Architecture: Agent architectures formalize agents through Profile, Memory, Planning, and Action modules, while cognitive and decomposition frameworks structure reasoning as search or progressively harder subtasks.These frameworks distinguish agents from LLMs merely equipped with tools and organize planning across modular reasoning steps.
- Reliability, Verification, and Governance: 91% lower p95 latency and over 90% token cost savings distinguish Mem0 from full-context baselines, while persistent computer use creates stochastic reliability, verification, and security challenges.Repeated runs may fail because of underspecified tasks and environmental changes; credentials, files, permissions, channels, and memory also expand the attack surface.
- Tool Use: Tool use progresses from executable calls to grounded API use, trajectory-level control, and standardized infrastructures, moving beyond locally valid actions toward coherent tool-use policies.The transition is not simply from fewer tools to more tools; it changes how actions are grounded, coordinated, and standardized.
- Agent-era Bottlenecks: The Agent era is constrained by ephemeral tool invocation, environmental brittleness, and absent long-term task closure, leaving artifacts transient and work products unreliable.Agents lack persistent state, recovery mechanisms, and verification loops for dependable end-to-end completion.
- OpenClaw Era: OpenClaw-style systems make a persistent managed workspace the organizing abstraction, preserving files, sessions, logs, tools, permissions, instructions, and skills across trajectories.Workspace operations can be inspected, validated, rolled back, and governed, shifting the conceptual unit from an attempted sequence to a persistent environment.
- Workspace Intelligence and Skills: Reusable skills enable composable capability packages and task closure, but they can become stale, over-specific, incompatible, or malicious, requiring metadata, dependencies, and versioning controls.Skill modularity changes the trust model by combining reusable domain expertise with supply-chain and compatibility risks.
4 Part III: Why “Workspace + Skill” Is the Key Leap
The next qualitative leap in agentic systems comes from combining reusable skills with persistent workspaces, turning isolated tool use into durable, verifiable task delivery. This paradigm supports delegation, task closure, and accountable digital work, while introducing brittleness, security risks, and substantial governance overhead.
- Persistent environments for durable task delivery: Persistent workspaces preserve files, terminals, logs, repositories, and execution contexts, allowing agents to accumulate state instead of reconstructing context after every tool call.They also determine visible state, executable operations, persistent artifacts, and auditable traces.
- Delegation and workflow: Workspace + Skill systems shift human–AI interaction from specifying micro-steps to delegating bounded objectives with constraints, permissions, success criteria, and acceptable risk.The paradigm imports professional workflow structure, including process, accountability, and completion criteria.
- Reusable skills: Skills externalize procedural knowledge into modular packages that can be tested, versioned, reused, parameterized, composed, refined, and inspected.Executable scripts and templates can reduce dependence on fragile natural-language reasoning.
- Workspace + Skill integration: Together, workspaces provide state and artifacts while skills provide procedures, constraints, and verification logic, enabling stronger task closure than either provides alone.OpenClaw exemplifies this architecture by combining a persistent local environment, skill directories, tool integrations, and task-oriented execution.
- Limitations of the Workspace + Skill paradigm: The paradigm raises the ceiling of agentic work but remains incomplete because persistent environments and reusable procedures introduce new failure modes.The limitations occur at the levels of skills, workspaces, and their interaction.
- Governance and evaluation: Reliable deployment requires controls for skill brittleness, supply-chain attacks, permissions, sandboxing, logging, reproducible state, testing, verification, and failure recovery.Evaluation becomes more expensive because success must be judged over trajectories and final workspace states rather than isolated responses.
5 Part IV: Data & Evaluation — Paradigm Shifts Behind the Scenes
Data and evaluation must shift alongside LLMs as they move from conversational chatbots toward reasoning systems, autonomous agents, and persistent OpenClaw-style workstations. Training and evaluation therefore evolve from static labels and prompt–response pairs toward verifiable trajectories, state–action–observation records, and evidence of task closure.
- Data and Evaluation: Data and evaluation jointly shape what model generations learn and how the scientific community defines progress.They function as dual pillars of AI development.
- Data and Evaluation: Training and evaluation must undergo a substantial paradigm shift as LLMs advance from conversational chatbots to reasoning systems, autonomous agents, and persistent workstations.The shift follows changes in both required training signals and evaluation methodologies.
- Data and Evaluation: Agentic and OpenClaw-style systems require reasoning traces and state–action–observation trajectories, including tool outputs, UI states, workspace changes, and final-state evidence.This replaces reliance on static prompt–response pairs and labels alone.
5.1 Data Paradigm Shift: From “Knowledge Corpus” to “Action Trajectory”
The data paradigm shifts from static knowledge corpora and instruction–answer pairs toward reasoning paths and interactive state–action–observation trajectories. In agent settings, reusable skills and verifiable workspace outcomes become part of the data, making collection substantially more demanding.
- From Knowledge Corpus to Action Trajectory: Training data progress from static knowledge corpora and instruction–answer pairs to reasoning-process data and, finally, action trajectories.The chatbot, Thinking LLM, and Agent/OpenClaw stages successively supervise answers, reasoning paths, and interactive work.
- Reasoning-Process Data: Thinking LLM supervision expands beyond final answers to Chain-of-Thought traces, intermediate steps, revisions, and verification signals.Process supervision scores individual reasoning steps rather than only whole-solution outcomes.
- Interactive Action Trajectories: Agent data record interactive sequences in which models observe environment states, select tool or UI actions, receive feedback, and continue toward task completion.Trajectories can include tool calls, multimodal UI actions, screenshots, DOM states, terminal output, and file changes.
- Reusable Skill Assets: Reusable skill assets become data, including operation recordings, command sequences, checklists, recovery procedures, and written experience.Voyager uses a retrievable executable skill library, while Agent S reuses external knowledge and internal experience through hierarchical planning.
- Data Scarcity and Verification: High-quality action trajectories are scarce because they require realistic tasks, configured workspaces, correct actions, and verifiable final states.WebArena, OSWorld, and ClawsBench expose this cost through executable environments, initial-state configuration, and final-state checks.
5.2 Evaluation Paradigm Shift: From Output Scoring to Task-State Verification
Evaluation progresses from scoring isolated final answers to judging reasoning processes and ultimately verifying whether agents achieve the intended closed workspace state. This shift adds requirements for reliability, safety, reproducibility, and auditable state changes.
- Stage I: Final-Output Evaluation: Stage I scores static-input tasks by final-output correctness, using metrics such as accuracy when labels or reference answers are clear.This stage is easy to scale automatically but cannot explain how the answer was obtained.
- Stage II: Process Judgment: Stage II evaluates long reasoning trajectories for coherence, faithfulness, correctness, and incorrect intermediate steps rather than final answers alone.LLM-as-a-judge and process-level verifiers inspect reasoning chains when simple reference labels are insufficient.
- Stage III: Task Closure: Stage III judges task closure: whether tool use and environment changes leave the workspace in the intended completed state.A valid reasoning trace is insufficient if code fails tests, an order is not submitted, a document is not updated, or a calendar event violates constraints.
- Workspace and OpenClaw Evaluation: OpenClaw-oriented benchmarks combine end-to-end success, reliability, reproducibility, and safety measurements across online tasks and simulated workspaces.ClawsBench evaluates productivity agents using services such as Gmail, Slack, Calendar, Docs, and Drive under reproducible state management.
- Auditability and Safety: Task-closure evaluation requires reproducible initial states, trajectory logs, replayable actions, and final-state diffs to support systematic auditing and comparison.Workspace evaluation also treats prompt injection, unintended operations, malicious skills, privacy leakage, and weak runtime protection as explicit risks.
6 Open Challenges and Future Directions
The paper frames the next challenge as making increasingly autonomous AI systems reliable in practice, because failures now extend beyond incorrect text to unsafe actions, corrupted states, incomplete closure, and untraceable long-horizon behavior. It therefore calls for integrated self-evolving ecosystems combining models, contexts, tools, skills, workspaces, and governance.
- From Model-Centric Capability to Ecosystem-Level Reliability: AI is progressing from language generation and question answering toward reasoning, tool use, and workspace-level task execution within digital environments.
- From Model-Centric Capability to Ecosystem-Level Reliability: Greater autonomy broadens failure modes to include unsafe tool calls, corrupted workspace states, incomplete task closure, and untraceable long-horizon behavior.
- From Model-Centric Capability to Ecosystem-Level Reliability: Future directions target self-evolving AI ecosystems that engineer models, contexts, tools, skills, workspaces, and governance mechanisms as an integrated whole.
6.1 Open Challenges: Making Autonomy Reliable
Reliable autonomy remains an open challenge because production agents must sustain task completion over long horizons while safely modifying stateful workspaces under permission, memory, and governance constraints. Addressing this challenge requires verification, recovery, persistent memory, operational controls, and socio-technical accountability that preserve human agency and data boundaries.
- From Impressive Demonstrations to Dependable Digital Work: Production agents remain far from trustworthy digital workers because isolated demonstrations do not establish stable long-horizon performance, safe constrained operation, persistent memory, or responsible social and organizational effects.The reliability gap extends beyond capability demonstrations to deployment conditions requiring stable performance, permission constraints, memory robustness, and consequence management.
- From Demonstrated Capability to Stable Completion: Long-horizon tasks can propagate tool-call errors, leave inconsistent workspace states after partial failures, and expose planning mistakes only after irreversible actions.Reliable autonomy therefore requires progress monitoring, intermediate verification, self-healing, and recovery mechanisms, especially when composing reusable skills into longer workflows.
- From Textual Guardrails to Operational Control: As agents access files, APIs, terminals, databases, and enterprise applications, safety requires permission isolation, risk-aware validation, audit trails, rollback, and inspectable human oversight.Agentic failures can include private-data leakage, file overwrites, external side effects, untrusted skill execution, and unauthorized decisions.
- From Technical Reliability to Socio-Technical Accountability: Reliable deployment must preserve meaningful human agency, attribution, contestability, escalation paths, and data-sovereignty boundaries as agents reshape professional work and process sensitive enterprise traces.Workspace traces may become memories, skills, evaluation examples, or training data, complicating boundaries among user data, enterprise assets, third-party information, and public system experience.
- From Short Interactions to Persistent Collaboration: Persistent collaboration requires memory of goals, constraints, decisions, failures, and environmental changes across sessions, tools, files, and users, while ultra-long contexts remain expensive and difficult to search.Memory must support working, factual, experiential, and procedural functions rather than relying only on ephemeral context.
6.2 Future Directions: Toward Self-Evolving AI Ecosystems
Future AI progress is framed as scaling self-evolving ecosystems around models, where workspaces, structured experience, reusable skills, verification, and governance convert actions into sustained system improvement. This trajectory moves AI from reactive chatbots toward adaptive digital colleagues that participate in workflows and accumulate experience.
- Beyond-Gradient Learning and System Stewardship: Self-evolving systems transform trajectories, workflows, failures, and tool errors into memories, skills, regression tests, wrapper updates, and permission rules.The research challenge is making these transformations reliable through curation, verification, lifecycle management, and human approval where needed.
- Digital Workspaces as Learning Environments: Structured workspaces make cognition actionable by preserving state, consequences, evidence, replay, rollback, permissions, provenance, and final-state differences.These primitives support both safety and learning because failures must be reconstructed and successes replayed before they can improve systems.
- Multi-Agent Orchestration and Governance: Multi-agent ecosystems require orchestration and governance to assign roles, decompose tasks, route artifacts, constrain authority, and prevent unsafe delegation or responsibility diffusion.Without coordination rules, collaboration can degenerate into parallel hallucination.
- From Chatbots to Digital Colleagues: Digital colleagues differ from chatbots by retaining project memory, learning local conventions, identifying recurring failures, maintaining shared tools, and adapting through experience.Validated memories, skills, workflows, tests, and policies enable sustained participation in collective problem solving.
- From Larger Models to Integrated Learning Ecosystems: Future progress shifts from scaling isolated models to scaling ecosystems that connect agents with workspaces, memories, tools, skills, evaluators, and governance.The model remains the cognitive core, while surrounding layers determine whether experience can accumulate.
7 Related Work
Prior work has advanced LLM reasoning, agent architectures, tool use, memory, skill accumulation, and interactive evaluation. This section distinguishes that foundation from the paper’s focus on operational reasoning with persistent states, verifiable workspace changes, and the Agent-to-OpenClaw boundary.
- Reasoning and self-improvement: Chain-of-Thought and Tree/Graph-of-Thought prompting showed that intermediate reasoning improves multi-step problem solving.Recent research extends this direction through inference-time scaling, process supervision, reinforcement-learning-driven reasoning, self-improvement, and reflection feedback loops.
- Reasoning and self-improvement: Unlike text-centered reasoning literature, this work studies reasoning integrated with tools, persistent states, and verifiable workspace changes.The paper frames operational reasoning as the point where internal capabilities become actionable in persistent environments.
- Agents, tools, and skills: Related agent research covers planning, memory, reliability safeguards, cognitive frameworks, external action steps, API invocation, task routing, tool utilization, and open-ended skill accumulation.Broad agent surveys are complemented by specialized studies of the components required for long-horizon behavior.
- Interactive evaluation: Interactive benchmarks evaluate agent abilities and assistant-style problem solving through environments for web navigation, shopping, and realistic website interaction.WebGPT connected language models with web navigation and human feedback, while WebShop, Mind2Mind, and WebArena test interactive web tasks.
- Research gap: Despite extensive related work, limited discussion addresses the boundary between the Agent Era and the OpenClaw Era.Existing studies supply ingredients such as external action loops, web and desktop environments, software repositories, stateful tools, execution feedback, memory mechanisms, and reflection loops.
8 Conclusion
The paper frames the shift from Chatbot to Digital Colleague as a transition from conversational answers to persistent work, combining deliberate Thinking LLM cognition with governed workstation execution.
- 8 Conclusion: The transition pairs next-token “fast thinking” with Thinking LLMs that leverage inference-time computation, and ad hoc tool-calling with OpenClaw workstation systems.OpenClaw systems add persistent workspaces, skills, and governance to task execution.
- 8 Conclusion: The “Workspace + Skill” paradigm enables persistent state and reusable procedures as mechanisms for shifting from episodic tool use toward persistent work.The supplied conclusion identifies state persistence and reusable procedures as drivers of the transition.