Network-AI: Framework-agnostic shared blackboard for multi-agent coordination

[Jovancoding]

Built an open-source TypeScript library that adds a shared blackboard (concurrent state store) to multi-agent systems — with atomic commits, budget tracking, and token-scoped security.

Why it exists:
When orchestrating agents across different frameworks, coordination primitives are usually locked to one ecosystem. Network-AI provides a neutral layer that any agent can read/write through, regardless of whether it's running in Agent Squad, AutoGen, LangChain, CrewAI, or raw function calls.

How it works with Agent Squad:
Agent Squad handles intent classification and routing beautifully. Network-AI slots in as a shared memory layer between your specialist agents:

import { LockedBlackboard } from 'network-ai';

const board = new LockedBlackboard();

// Research agent writes findings
const id = board.proposeChange('researcher', {
field: 'context.findings',
value: searchResults
});
board.validateChange(id);
board.commitChange(id);

// Writer agent reads them — no message passing needed
const findings = board.read('context.findings');

Each agent gets scoped access via AuthGuardian tokens — so your classifier agent can't accidentally overwrite data owned by a specialist.

Key features:

Atomic commit workflow (propose → validate → commit) with automatic rollback
Per-agent budget tracking — cap token/cost spend per agent
AuthGuardian: HMAC-signed tokens with trust levels and scope restrictions
Adapters for 12 frameworks (including a generic one for Agent Squad integration)
251 tests, zero external dependencies
Path-traversal hardened (regex + resolved-path boundary checks)
Links:

GitHub: github.com/jovanSAPFIONEER/Network-AI
npm: npm install network-ai
Docs: README
Would love feedback from anyone running multi-agent setups with Agent Squad — especially around how you currently handle shared state between specialist agents. Happy to add a first-class Agent Squad adapter if there's interest.