Share Context, Not Data: Governed Semantic Abstraction for Agent-to-Agent Context Sharing
AI agents increasingly act for people in dealings with other parties' agents. To be useful, an agent must tell the other agent something about the person it acts for. Today the limit on what it tells is usually an instruction to a language model that holds the person's records. In this position paper, we argue that agents should exchange a derived, typed, citable abstraction of a person's records, not the records themselves, and that deterministic gates, not a model's judgement, should enforce what is disclosed. Every record admits a lattice of true, progressively less specific statements. The person fixes a level of abstraction per recipient and purpose in a contract, and can revoke any grant. A language model performs the abstraction, and gates maintain three invariants: what the ontology excludes does not exist in the derived model, what a contract does not grant is not retrieved, and no claim is delivered unless its citations resolve to retrieved facts. Every answer is logged for the person it describes. We state which adversaries the design does and does not defeat, respond to five objections, and set out open problems. The first is to bound what a recipient can still infer from a well-typed abstraction.
Authors
- Daniel Platnick (ORCID: https://orcid.org/0009-0009-1262-9248)
- Marjan Alirezaie (ORCID: https://orcid.org/0000-0002-4001-2087)
- Hossein Rahnama
- Sourena Khanzadeh
Institutions
- Toronto Metropolitan University (CA)
- Massachusetts Institute of Technology (US)
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-20
- DOI
- https://doi.org/10.5281/zenodo.24407188
- Primary Topic
- Access Control and Trust
- Type
- article
- Field-Weighted Citation Impact
- 0.00