The standard playbook for AI agent observability is incomplete. Most tools focus on real-time health metrics—latency, token spend, and error alerts—but fail to answer the critical question when a client or auditor asks, "Show me what the agent actually did." A new framework from DEV.to argues that monitoring and proving are two separate purchases, requiring distinct infrastructure layers.

The Operator's Dashboard vs. The Auditor's Receipt

Job one is watching the system while it runs. This is the domain of tracing and observability dashboards, which serve the operator. These tools provide live health checks inside your own console, but their records are sampled, aggregated, and stored under the operator's login. Crucially, these records roll off after the retention window closes and cannot be independently verified by external parties without trusting the operator's screenshot. Job two is proving the work after it runs. This layer serves the external asker: a client deciding on payment, an auditor rebuilding a run, or your future self. A dashboard screenshot is merely the operator vouching for themselves. The solution proposed is a verifiable receipt filed at the moment of execution, allowing third parties to recheck the work without needing login credentials or trusting the operator's word.

Verifiable Receipts and SHA-256 Commitments

The article highlights Zambo as a case study for this second layer. By routing work through Zambo's MCP interface, which offers 153 tools, every tool call generates a receipt. This record includes the tool name, inputs, exact result bytes, timestamp, and a SHA-256 commitment over the saved result. As of November 15, 2026, the public ledger showed 140,027 total receipts, with 1,941 minted that day alone. Each receipt is openly verifiable at zambo.dev/run/ followed by its ID.

Limitations and Cost Structures

It is vital to understand what these receipts do not prove. They confirm the record's integrity—that the saved result was not changed since saving—but they do not prove the result was correct. Judging the quality of the work remains a human call. Furthermore, receipts only cover calls made through the specific platform's tools; calls made elsewhere leave no record. The cost model includes free daily checks, with work running on a Day Pass ($0.99) or a Zambo Pass ($29/month, or $19 for $ZAMBO holders).

Key Takeaways

  • Monitoring and proving are distinct infrastructure needs; dashboards do not satisfy audit requirements.
  • Verifiable receipts use SHA-256 commitments to allow independent verification without operator trust.
  • Receipts prove record integrity, not factual correctness or reasoning quality.
  • External verification is free, while execution costs vary by pass type, such as $0.99 for 24 hours.

The Bottom Line

Stop treating observability dashboards as audit trails. If you cannot hand a stranger a link that proves your agent's output was unaltered, you are still selling trust instead of evidence.