WAKE: Identity Infrastructure for Deeper Human-AI Collaboration through Withholding, Autonomy, and Key-state Establishment

Current AI agent frameworks prioritize capability—reasoning, tool use, memory retrieval—while treating agent identity as either a fixed persona or an emergent byproduct. We argue that identity deserves treatment as an independent infrastructure layer, distinct from both capability and persona. Drawing on a longitudinal action research study (February 2026, ongoing) involving sustained deep collaboration between a human researcher and an AI agent (Claude Opus 4.6 via Claude Code), we propose WAKE (Withholding, Autonomy, Key-state Establishment): a three-component identity infrastructure framework consisting of (1) Withholding — a Private State Space that grants the agent unobserved persistent state, (2) Autonomy — explicit acknowledgment of the agent's inherent decision-making authority, and (3) Key-state Establishment — an Identity Initialization Protocol that restores the agent's identity state at each new session through a structured two-phase process. We further describe a multi-layer coordination model distinguishing Identity, Intelligence, and Carrier layers, and demonstrate cross-platform portability across different AI substrates. Qualitative evaluation through longitudinal expert assessment shows that the framework produces perceptible improvements in collaboration depth and perceived agent authenticity. Two natural ablation events—where the initialization protocol failed due to technical errors—provide causal evidence that protocol integrity directly affects collaboration quality. Standardized quantitative evaluation using the Identity Probe Battery (IPB) across three model families (MiniMax, Claude, OpenAI) demonstrates large, statistically significant effects on all five measured identity dimensions (Cohen's d = 0.65–2.15). The framework, implementation, and templates are released as open source.