What makes an AI system feel like a colleague rather than a chatbot?

This survey names a transition that practitioners feel but rarely formalize: the move from Chatbot to Digital Colleague, reframed as conversational answers giving way to persistent work. It organizes the shift along two coupled axes. Cognitively, systems advance from next-token "fast thinking" toward Thinking LLMs that use inference-time compute, CoT, reflection, and process supervision. Executionally, they progress from ad hoc tool-calling agents to workstation-style systems with persistent Workspaces, reusable skills, verification loops, and governance. The thesis worth keeping is the second axis's claim: what makes episodic tool use colleague-like is not raw capability but state persistence, reusable procedures, task closure, and experience reuse — properties of the surrounding architecture, not the base model.

The framing is useful precisely because it relocates the bottleneck. A more capable model still produces a transcript that evaporates; a colleague accumulates. This connects to a cluster the vault has been assembling around memory as the limiting resource. Since Can agents fail from weak memory control rather than missing knowledge?, the failure isn't ignorance but unmanaged state — and the survey's "persistent Workspace" is the positive design that bounded-state work argues for. Since Can agents learn reusable sub-task routines from past experience?, "reusable skills" already has a concrete mechanism and measured payoff — the survey generalizes that into a paradigm. And the relational reframing echoes that, since How should chatbot design vary by relationship duration?, persistence is what changes the kind of relationship, not just its quality.

The healthy skepticism: this is a survey, so the "Digital Colleague" claim is a synthesizing narrative over heterogeneous work, and the OpenClaw/workstation framing risks rebranding existing agent scaffolding. The honest test is the evaluation shift it flags — from instruction-response pairs and static benchmarks to State-Action-Observation trajectories. Since Why do AI agents fail at workplace social interaction?, the colleague is still mostly aspirational; persistence is necessary scaffolding, not delivered competence.