Skip to main content
Early Preview. Agent Tasks is in early development and not yet generally available. Capabilities and setup will change before release.

​
Overview

Correlate a firing Honeybadger incident with recent changes, logs and metrics, then report and open a PR. When an incident fires, the agent investigates it and hands a clear summary back to the human on-call, it never merges a fix on its own.

​
How It Works

When an incident fires, the agent:
  1. Reads the incident details (id, title, severity, affected services) from the trigger, fetching any missing details from Honeybadger.
  2. Correlates the incident with recent change context: the latest deployments, config changes, and merged PRs for the affected services around the incident start time.
  3. Pulls the relevant signals: application logs, metrics, and any runbooks for the affected services.
  4. Determines the most likely root cause and the blast radius, and is explicit about its confidence and what it could not verify.
  5. Reports its findings to the on-call human in chat: a short summary, the suspected root cause, and a recommended next step.
  6. If the fix is small and well-understood, opens a PR with the proposed change and links it in the message. It never merges, the human always stays the gate.

​
Setting It Up

1

Create the agent task

From the Agent use cases section when creating a new service, pick Incident Analyzer with Honeybadger (see Creating an Agent), or start from scratch and paste in the instructions above.
2

Connect the repository

Add the Git repository of the service(s) this agent should investigate as its Context, so it can read recent commits and merged PRs around the incident. You can also add the Qovery service(s) themselves as Context, in addition to or instead of their repository.
3

Add the Qovery MCP server

Add Qovery’s MCP server (https://mcp.qovery.com/mcp) so the agent has more context on the applications it needs to monitor, deployment history, logs, and metrics for the affected services. A read-only API Policy Token is enough, this agent only reads and reports, it never applies a fix directly.
4

Choose a Provider

Enter your Anthropic API key, the only provider supported today.
5

Add credentials and domain allowlist

Add your Honeybadger API key as a secret, referenced by name in the instructions above (for example HONEYBADGER_API_KEY), for more context on the incident if needed. In Honeybadger, go to Settings > API keys to create one. Add its hostname to the domain allowlist, for example api.honeybadger.io.
6

Add a trigger

Select From a webhook, the default for this template, and copy the generated webhook URL from the agent task’s overview page, you’ll need it next.
7

Point Honeybadger at the webhook

In Honeybadger, go to Alerts & Integrations, add the Webhook integration, and paste the URL from the previous step. Under Error events, check When an error is initially reported or when reported again after being resolved, then save.
Honeybadger webhook integration settings with the webhook URL and the error event set to when an error is initially reported or reported again after being resolved
8

Add an output (optional)

Same as any other agent task, see Outputs.
9

Use Clone Environment as the execution mode

Incidents can fire frequently, so use Clone Environment rather than In Place: an overlapping trigger would otherwise replace the currently running investigation. See Execution Mode.
10

Trigger a first run and check the results

Test the agent by clicking Trigger on its overview page, or by sending a POST request with a sample incident payload to its webhook URL. Check the run under the agent task’s Deployments tab, and confirm the summary reaches the Slack channel tied to your webhook URL.
Resources: this agent mainly makes API calls and reads logs and metrics, a small allocation is enough to start. See Configuration for how resources are set, and Triggers and Outputs for how triggers work in general.