The Control Plane for an Agentic Society

​Every gain in AI so far has been vertical: bigger models, deeper reasoning inside a single agent. That has a ceiling. Practical intelligence, the kind that screens a payment, closes a purchase or settles a position, is collective: agents that work together, delegate to one another, reason jointly and act. Cells did it. Traders did it. Firms did it. Software agents are doing it now.

​The catch is that collaboration and delegation only scale with two ingredients most agent frameworks still lack: a boundary on what each agent may see and do, and enforcement when it is crossed. Prompts and policies are neither. A society of agents needs a control plane: mandates enforced in the protocol, and workflows in which agents compute jointly on inputs none of them can see. Privacy is not a tax on the outcome here; it is what makes the outcome optimal. The moment an agent's private state can leak, it stops reporting truthfully and the joint result degrades.

​The discussion is anchored on AP3, the Agent Privacy-Preserving Protocol that Silence Laboratories open-sourced to LF Decentralized Trust at the DeCompute 26 San Francisco chapter: an open substrate for data-blind multi-agent workflows, with mandates as the boundary, cryptography as the enforcement, and Google's A2A and AP2 as the connectivity and settlement rails. We will use it as a reference design to test against the three cases below, not as the answer.

​This invite-only roundtable brings together senior leaders from banks, payment networks, fintechs and regtechs, alongside the LFDT community and agentic-security practitioners, under Chatham House rules.

​Themes

• ​Anti-fraud: Mule networks, synthetic identities and coordinated attacks that no single detection system sees; agents resolving entities and sharing signals across trust boundaries while raw records stay where they are.

• ​Agentic commerce: Agents transacting under mandates and spend limits on a principal's behalf, without revealing the principal's intent to the counterparty.

• ​Treasury and settlement: Agents negotiating liquidity, FX and routing while positions, limits and client flows stay private.

• ​The control plane: Consent, authority, delegation, revocation and auditability as agents cross trust boundaries, inside an institution and between them.

​Expected Outcomes

• ​A reference model for data-blind multi-agent workflows across the three cases: architecture, privacy, mandates, governance and incentives.

• ​A shortlist of high-value use cases ready for a pilot or consortium, with AP3 and LFDT as an open home for the work.

• ​A practitioner working group to take a priority use case to proof-of-concept, and an industry white paper drawing on the discussion.

​Who Should Attend

​Fraud, financial crime and AML leaders; heads of payments, treasury and digital assets; and the AI, security and privacy teams at banks, payment networks, fintechs and regtechs building agent-driven workflows.