In February 2026, we deployed 13 AI agents across 5 platforms. Telegram, Instagram, Facebook, web, and WhatsApp. Real agents doing real work: responding to leads, filing grants, managing legal docs, coaching users through 6,854 behavioral scenarios, and processing financial data.
By day 3, one agent tried to delete a production config file. By day 5, another started drifting from its assigned role. By week 2, we had caught 99+ policy violations across the fleet - none of which bypassed our governance layer.
We did not wait for a framework to tell us this was a problem. We built the solution ourselves.
What We Built (and When)
February 2026 - The Nervous System v1.0 went live. 7 enforced behavioral rules. A preflight check that blocks edits to 99 protected files. A SHA-256 hash-chained audit trail for every action. A kill switch that stops any agent in one command. Published to npm as mcp-nervous-system. Listed in the Anthropic MCP directory.
March 2026 - Tamara, our operations manager agent, started running autonomous health checks every 60 minutes. She dispatches agents, monitors for drift, and reports to the founder via Telegram. No human in the loop for routine operations.
March 13, 2026 - We deployed Claude Code hooks (PreToolUse, PostToolUse, Stop) that route every tool call through the Nervous System before execution. Safe commands pass. Dangerous commands get blocked in real time with an audit entry.
March 25, 2026 - NS Governor v2.1 went live. Stateful escalation: warn on first violation, strike on second, auto-kill on third. 29 policy rules. Sliding 15-minute window. Proven in a live demo where destructive commands were blocked, the agent was killed, and the target file survived.
March 26, 2026 - SQLite audit brain replaced flat JSON logging. Every decision now persists with full traceability: agent ID, session, tool, action, policy rule, risk category, escalation level. Queryable. Exportable.
Same day - YAML policy engine deployed. 24 policy files: 1 global default, 10 role-level policies, 13 agent-specific overrides. Policy changes take effect without code changes. Just edit the YAML and reload.
Why This Matters
Every week, a new agent framework launches. They all focus on the same thing: making agents smarter, faster, more capable.
Nobody asks: who controls them?
We did. Before it was a feature in any framework. Before governance was a checkbox on a product page. We built it because we needed it. 13 agents running 24/7 with access to production files, email accounts, financial data, and government portals. You do not run that without a control layer.
What We Have Now
- 13 governed agents across 5 platforms
- Governor v2.1 with stateful escalation and auto-kill
- SQLite persistent audit trail (every decision logged)
- 24 YAML policy files (global, role, agent-level)
- 30+ MCP tools published on npm
- SAM.gov registered, CAGE code assigned, government-ready
- Running on a single VPS for under $300/month
The Nervous System is open source: npmjs.com/package/mcp-nervous-system
The question was never whether AI agents would need governance. The question was whether anyone would build it before something went wrong.
We did.
Arthur Palyan is the founder of Levels Of Self, an AI governance and multi-agent infrastructure company based in California. The Nervous System MCP is listed in the Anthropic MCP directory and published on npm.
Top comments (0)