pk.Prashant Kumar
AI Architect · Author · Building POCs
Issue 01 · Vol. I Currently building — Avarna v4 · Stock Predictor · Secure Multi-Agent — · 2026
← All POCs
03Agents
Vision · early build · v0 in progress

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.

·Why
The gap

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.

·Vision
The vision

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.

·Now
Current status

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.