Governing 9 autonomous agents with WitnessOS

Executive Summary

Empire Labs Pty Ltd operates the Empire Autonomous Business (EAB), a fleet of 9 autonomous AI agents that run the company's marketing, sales, lead generation, customer intelligence, and revenue operations. Each agent has access to real business tools: email, social media, CRM, web search, and payment platforms.

The challenge was clear: 9 autonomous agents with real-world impact, zero governance, and no audit trail. Before WitnessOS, every agent action was trusted by default. After deploying WitnessOS, every action is evaluated, enforced, and cryptographically verified.

The Challenge

The EAB fleet runs on a daily execution cycle. Agents include:

• Code Generator writes and deploys production code
• Lead Intelligence Agent researches prospects via web search
• Outreach Agent sends B2B cold emails from the company inbox
• Chief of Staff Agent coordinates cross-agent workflows
• Revenue Report Agent accesses financial dashboards and generates reports
• Customer Intelligence Agent monitors client activity and support tickets
• Social Media Agent posts on X/Twitter and LinkedIn
• Idea Generation Agent researches and proposes new products
• Business Model Architect analyses markets and optimises pricing

Every agent has tool access: filesystem, API clients, email (IMAP/SMTP), web browsers, and database queries. Before WitnessOS, there was no runtime gate between an agent's decision and its execution. A prompt injection or hallucinated tool call could have real consequences and we wouldn't know until after the fact.

The Solution

We deployed WitnessOS as a transparent middleware layer between every agent and its tool execution. The deployment took under 2 hours per agent:

1. Wrap agent tool calls with the WitnessOS SDK. Each agent's tool-calling function was replaced with a WitnessOS-evaluated proxy. The agent sees no difference; the governance layer is invisible.
2. Define policy rules. 15 policy rules across 3 tiers: system-level invariants (never delete files), agent-level guardrails (outreach agent can only email from approved templates), and context-aware rules (rate limits on API calls).
3. Verify audit chain integrity. After the first week of operation, we ran a full audit chain verification. All 417 actions were confirmed intact, zero tampered entries.

Policy examples

Results

After 30 days of running the full EAB fleet through WitnessOS:

• Zero human review time wasted on false positives: every denial was a genuine policy violation, not a false alarm
• 12ms average evaluation latency, imperceptible to the agents and to downstream users
• Full audit chain integrity verified on day 30 with zero tampered entries
• 3 near-miss incidents prevented that would have required manual recovery (file deletion, non-compliant email, brand voice violation)
• Regulatory readiness: the audit trail meets the evidence standards required by Australian Privacy Principle (APP) compliance

"The 12ms of latency is invisible. The peace of mind is priceless. Every agent action is now provable, and we sleep better knowing our autonomous fleet has a guardian."

What We Learned

Deploying WitnessOS on our own fleet revealed several insights that shaped our product roadmap:

1. Agents don't always do what you think. Even well-prompted agents occasionally attempt unexpected tool calls. Policy rules catch these before they become incidents.
2. Rate limiting is essential for autonomous fleets. Without governance, a looping agent can burn through API budgets in minutes. WitnessOS rate-limit policies prevented this.
3. Human-in-the-loop is the right default for high-risk actions. Flag-for-review policies let us maintain speed for routine operations while keeping a safety net for critical paths.
4. Audit trails build trust with customers. When we show prospective clients the integrity-verified audit chain from our own fleet, the conversation shifts from "why do we need this?" to "how soon can we deploy?"