{{THEME_SCRIPT}} {{HEADER}}
← All prototypes

ROOT18D / EXPERIMENT 012

Break the response.
Keep the progress.

An agent’s hardest job is finishing what it started. Interrupt a workflow, recover its checkpoint, and see whether the action happens twice.

AGENT RECOVERY LABSave.
Interrupt.
Resume.Durable execution / interactive demo
Real saved checkpoints. Synthetic incidents and actions. Deterministic decision logic; no AI model is called.

02 / FOLLOW THE EXECUTION

Ready when you are.

Idle
Steps saved0 / 4
Remediations applied0
Checkpoint version—

Choose a scenario, then start. You will approve or deny the proposed remediation.

  1. 01Inspect
  2. 02Diagnose
  3. 03Remediate
  4. 04Verify

Execution journal

  1. Your execution history will appear here. Every entry comes from a server checkpoint.
Inspect saved state
No checkpoint yet.

03 / THE MECHANICS

A retry should never mean
“do everything again.”

The interesting failure happens after an action succeeds but before its response reaches the caller. This lab exposes that gap.

Explore the Agent Harness Starter Kit ↗
  1. One request. One durable transition.

    Each step saves its state to private Vercel Blob storage. Subsequent requests read that checkpoint without relying on server memory.

  2. Approval is part of the record.

    The server rejects remediation until you approve it. Denial ends the run without applying an action.

  3. Save the action and its receipt together.

    The simulated effect and checkpoint share a conditional write. Stale or concurrent requests cannot both advance the same version.

  4. Recover from evidence.

    The injected HTTP 503 happens after saving. Recovery reads the saved action receipt, skips remediation, and continues to verification.

What this demonstrates

This is a request-driven reference state machine, not a background worker. Pausing stops the browser from requesting the next step; it does not terminate a running server. The fault injects a lost response, not an operating-system crash. External APIs would need their own idempotency keys or an outbox; this demo’s single-write guarantee covers only the simulated action stored with its checkpoint.

All telemetry and health-check values are synthetic fixtures. No cloud account is modified. The exported trace excludes the private run key. Anyone holding that key can access or delete the run; this page stores it only in your browser and does not place it in the URL.

Built on Vercel Blob conditional writes ↗. Related reading: The Agent Harness ↗.

{{FOOTER}}