Secure Multi-Agent Framework.
A substrate for cooperating agents that treats security and privacy as load-bearing — not as a label slapped on at the demo stage.
Most agent frameworks pretend security is somebody else's job.
Open the docs for any popular multi-agent framework and you'll find lots about planning, tool-use, memory, and inter-agent messaging. You'll find very little about how to stop one agent from leaking the secrets of another. Almost nothing about how to enforce per-agent data access policies. Effectively nothing about audit, lineage, or "which agent saw what, when, and under whose authority."
That's fine in a demo. It is not fine the first time a multi-agent system gets near healthcare records, financial data, or anything covered by a real regulation. The right time to design for that is at the framework layer, not in a frantic retrofit six months after launch.
A substrate, not another wrapper.
The aim is a framework where every agent has an identity, every tool call is logged with provenance, every piece of context an agent receives carries a data-classification tag, and every cross-agent handoff is subject to a policy check. None of that should be opt-in. All of it should compose naturally with how developers actually build agents.
Concretely: agent identity that's verifiable, capability-scoped tools instead of unbounded function calls, structured memory with classification labels, an audit log that's queryable rather than a wall of JSON, and a policy engine that compliance teams can read without learning Python.
The natural integration with Avarna is part of the point: privacy enforcement at the substrate level, not in app code.
Where it is today.
Design document in progress. Core abstractions sketched on paper (agent, capability, context tag, audit event, policy). A v0 reference implementation is starting to take shape — small, narrow, opinionated. The first proper writeup lands on this page when there's something runnable worth showing.
If you work on multi-agent systems and you've felt this same gap, or you've been forced to retrofit security into a framework that didn't plan for it, I want to compare notes. Email's at the bottom.