---
title: "Infrastructure and evidence"
description: "Observability discovery, local receipt verification, and the next infrastructure steps."
canonical: "https://agenttwin.si/docs/infrastructure/"
---

# Infrastructure and evidence

Observability discovery, local receipt verification, and the next infrastructure steps.


AgentTwin maps the services your application depends on and records what a workflow test did. Infrastructure monitoring and test-world evidence serve different purposes: monitoring explains how AgentTwin operates, while test evidence explains the result of a specific workflow.

## Observability in your environment

Discovery recognizes Better Stack and OpenTelemetry SDKs, imports and configuration names. Compose, Dockerfile and Kubernetes manifests can identify Better Stack collector, eBPF collector and standard OpenTelemetry collector images. These resources use the existing inventory and application relationship machinery.

Detection establishes configuration evidence. It does not establish that a collector is running, show production traffic, or provide a service twin. Better Stack and OpenTelemetry currently have no implemented twin or network probe.

## Verify a local receipt

After a local run, check its recorded evidence against the current world specification and application files:

```sh
agenttwin verify --file .agenttwin/runs/RUN_ID/run.json --world agenttwin.world.json --json
```

Replace `RUN_ID` with your run ID. Application paths are resolved relative to the world specification. Verification reads files and evidence; it does not rerun your application or call a model.

The verifier checks specification and source digests, the exact staged file set, the application result, seeded/final state coverage, recalculated HTTP state assertions, unsupported operations and requested faults.

| Verdict | Meaning | Exit code |
|---|---|---|
| verified | Current files match and recorded checks pass | 0 |
| stale | Source files or specification changed, or source evidence is missing | 2 |
| failed | Recorded execution/state checks fail | 1 |
| invalid | Receipt or specification is invalid, inconsistent or outside the supported runtime | 1 |

This command supports local CLI receipts. Hosted/E2B receipts and database assertions are outside its scope. A verified receipt remains self-reported evidence; it is not independent execution attestation. Digests cover explicitly staged files, not every dependency or external service. The command does not update dashboard status or automatically gate sync.

## Infrastructure direction

The next operational work is correlated telemetry for API requests, queue wait, execution duration and sandbox cleanup, followed by a durable store and queue before scaling replicas. Future scenario evaluation should separately check required outcomes and forbidden side effects.

Useful references are the [Better Stack collector](https://github.com/BetterStackHQ/collector), [Claude Referee](https://github.com/ismaildasci/claude-referee), and [ArgaBench](https://github.com/argalabs/arga-twins-benchmark). AgentTwin's changes are original implementations. No external collector or evaluation service was installed by this work.