An LLM agent needs to plan actions in your environment. Your environment contains source code, credentials, file paths. The LLM runs in the cloud. The planning requires understanding the environment. The environment cannot be sent to the cloud.
Note (2026-11-12; read against the paper's full text): the paper defines (k,delta)-anonymity and epsilon-unlinkability as architectural privacy goals; it describes the cumulative disclosure budget as a simple weighted field-cost rule that is not presented as a full information-theoretic guarantee, and it says the non-disclosure guarantee comes from the architecture, which strips raw content before it reaches the cloud. The essay's statements that the guarantees are formal and that the budget ensures multiple interactions cannot compose into a reconstruction are stronger than that. The results (SND of 1.0, plan quality above 0.79 for three of the four planners) are from a benchmark of 60 synthetic tasks in ten domains; the fourth planner, Gemini 3 Flash, reached 0.689, and with Kimi K2.5 the plan quality was 0.808 against 0.843 for raw context.
Yu et al. build PlanTwin: a middleware that creates a sanitized digital twin of the environment — an abstraction that preserves everything the planner needs to know while removing everything that could be reconstructed into sensitive information. The twin is schema-constrained: it maps the real environment into a planning-oriented representation where structure is preserved but content is replaced.
The privacy guarantees are formal. (k,δ)-anonymity ensures that sensitive items cannot be uniquely identified. ε-unlinkability ensures that multiple interactions do not compose into a reconstruction attack. The compositional guarantee matters — a planner that sees one abstracted view per query might reconstruct the original across many queries, and PlanTwin's disclosure budget tracks cumulative leakage.
The results: full sensitive-item non-disclosure (SND = 1.0) with planning quality exceeding 0.79 for three of four planners tested. The abstraction costs something — not every planning detail survives sanitization — but the degradation is small relative to the privacy gained.
The structural tension: planning requires understanding, and understanding requires information, and information leaks. The twin resolves this by distinguishing between structural information (which the planner needs) and content information (which the planner must not see). The same distinction that makes abstraction useful in programming — hiding implementation behind interface — makes it useful in privacy: hiding sensitive content behind planning-relevant structure.