> ## Documentation Index
> Fetch the complete documentation index at: https://www.voiceflow.com/docs/llms.txt
> Use this file to discover all available pages before exploring further.

# Investigate a failed automation

> Reconcile the original execution before retrying a request.

Start with a failed, incomplete or timed-out tool call and access to the provider's execution history. The workflow owner must be able to inspect the destination system; the agent transcript alone cannot establish whether a write happened.

## Trace the original request

### 1. Identify the exact call

In **Run** or the failed conversation's [Transcript](/docs/documentation/measure/transcripts), record the tool, time, request ID and target Environment/version. Keep credential values out of the investigation notes. Distinguish the attached agent tool from a standalone Draft execution.

### 2. Check configuration before replaying

Open the provider manual and compare the deployed endpoint/scenario, authentication, input names/types and response contract. Confirm the selected credential slot or Secret override. For MCP, refresh and inspect the current tool schema after an upstream change. Listing a tool does not prove its business action is authorized.

### 3. Find the provider execution and destination record

Use Make scenario history, Zap History/MCP History, n8n Executions, Power Automate run history, the Workato recipe job, the Boomi process execution or the Mule application logs. Ask the workflow owner to match the original request and inspect every required side effect. A later step can fail after a record was already created.

### 4. Choose the recovery from the observed state

If the intended record exists, use it and repair only the remaining work. If no work occurred and the failure is fixed, retry according to the provider's contract. If the result is unknown, keep it unconfirmed and investigate; do not generate another request ID to bypass deduplication. For an intentional new request, collect confirmation and create a new ID through the application.

## Read the result correctly

| Observation | Interpretation and next action |
| --- | --- |
| 401/403 | Check the actual caller/credential and its resource access. A connected account or catalog listing does not prove action permission. |
| 404/405 | Verify the deployed URL, active listener/workflow and HTTP method. |
| 2xx with an application error or missing record | Inspect the provider body and downstream state; do not announce business success. |
| Pending/queued receipt | Follow the existing status/notification route; the work is still pending. |
| Timeout after a write | Reconcile the original request before retrying. |
| An older captured value remains | Clear the result variables and check the current call. A missing capture does not necessarily erase the previous value. |

For standalone API-tool calls, the business body is under `output.body`. Attached API tools and Workflow capture paths use the provider body directly. MCP and packaged integration results have their own envelopes: use the provider manual and actual response rather than copying an API-tool path.

## Verify the repaired journey

Repeat a safe permitted request and the original failure case. Inspect the current call, final record and agent's wording, then test the intended published client after a credential/configuration change. Confirm that no duplicate record was created. Keep source/configuration checks distinct from a successful authenticated provider execution.
