# Prerequisites and Troubleshooting

---

## What a Task needs

Four things. **Create Task** checks them for you and names anything missing, with a link to fix it.

| You need                           | Notes                                                                                                                         |
| ---------------------------------- | ----------------------------------------------------------------------------------------------------------------------------- |
| A runner in your AWS account       | Tasks run on your own infrastructure. See [Runner Installation](/docs/runner-installation/architecture/).                     |
| It to be online                    | If you've nominated a default runner, that's the one that has to be up.                                                       |
| An LLM key                         | See [LLM Providers](/docs/ai-providers/llm-providers/).                                                                       |
| Repositories the agent can push to | It needs permission to push branches and open pull requests. See [Adding a Repository](/docs/dashboard/adding-a-repository/). |

**Which provider?** It follows the agent and the model: Codex needs your OpenAI key; Claude Code needs Anthropic — an API key or [AWS Bedrock](/docs/ai-providers/aws-bedrock/); opencode uses whichever provider serves the model you picked, so Anthropic or Bedrock for Claude models and OpenAI for GPT models.

## Common problems

**"Your runner isn't active."**
It hasn't checked in recently. Start it, give it a moment, then hit **Refresh**.

**The Task gets all the way through and then fails.**
Almost always permissions: the agent wrote the code but couldn't push it or open the pull request. Reconnect the repository with permission to write code and open pull requests — installations set up before Tasks existed often only have read access, which is why it fails at the very end rather than the start.

**It failed its own checks.**
Your build or tests didn't pass, so the agent committed nothing. That's deliberate. Read the log to see what broke, then re-run with feedback pointing at it.

**The agent couldn't reach something.**
Agents run with restricted network access. If it was blocked, you'll see which host in both the run details and the pull request.

Add it under **Organization Settings → Agents → Allowed Hosts** and re-run. Hostnames must be exact — no wildcards or IP addresses. Usually this is a private package registry or an internal API.

**It's sitting there not starting.**
Normal if another Task is running — they go one at a time, oldest first. If nothing else is running, check your runner is online.

**Nothing happened and there's no pull request.**
Check whether the repositories show **No changes** — the agent may have concluded nothing needed doing. Its summary will say why.

## Who can do what

Tasks follow your [roles and permissions](/docs/role-based-access-controls/introduction/). Creating, viewing, starting and deleting are separate, so a teammate might be able to see Tasks without being able to start one.

---

Source: https://deployment.io/docs/tasks/prerequisites-and-troubleshooting/