# Troubleshooting

> The agent did the wrong thing, the run will not finish, nothing happened — the common symptoms and what actually fixes each one.

Symptoms, in the order they come up. Connector-specific problems have their [own page](/docs/connect/troubleshooting/).

## The agent did the wrong thing

Read the [run](/docs/work/runs/) before changing anything — the tool calls show where it went wrong, and the fix follows from that.

| What you see in the run | Cause | Fix |
| --- | --- | --- |
| It called no tools and answered from general knowledge | It does not know it should look something up | Say so in the [system prompt](/docs/build/system-prompt/); check the connector is bound |
| It called the right tool with wrong arguments | Missing context | Put the fact in the prompt or [memory](/docs/capabilities/memory/) |
| It looped without converging | Task underspecified, or [effort](/docs/build/models/#reasoning-effort) too low for its length | Narrow the task; raise effort |
| The answer is right but unusable | No stated standard for "done" | Describe the output you want in the prompt |
| It asked you something obvious | The fact is not anywhere it can see | Prompt, memory, or [project](/docs/capabilities/projects/) |

If one steering sentence reliably rescues the run, that sentence belongs in the prompt. See [Testing and iterating](/docs/build/iterating/).

## The run is not finishing

**It is paused, waiting for you.** A [question](/docs/work/questions/), an [approval](/docs/work/approvals/), or a [skill request](/docs/capabilities/skills/#when-an-agent-asks-for-a-skill) holds the run indefinitely. Scroll to the bottom of the conversation — the prompt is there.

**It hit the step limit.** Long tasks end without a final answer. Ask it to continue, or split the task.

**It failed.** The error is in the conversation. A model error usually means the model provider rejected the call — with [BYOK](/docs/account/bring-your-own-key/), check the key and its credit.

**You closed the app.** That does not stop anything. Runs finish on Endue's servers and the result is in the thread when you return.

## Nothing happens when I send a message

- **A run is already going.** A message sent during a run [steers](/docs/work/steering/) it rather than starting a new turn.
- **You are out of allowance.** Runs are declined when the week's allowance is exhausted — see [Plans and usage](/docs/account/plans-and-usage/).
- **You were signed out.** Sessions end after a period of inactivity; sign in again and the conversation is where you left it.

## The agent lost a capability it had

| Symptom | Check |
| --- | --- |
| A connector's tools are gone | The connection needs re-authorizing, or was unbound — see [Troubleshooting connections](/docs/connect/troubleshooting/) |
| It stopped remembering things | The memory [tool group](/docs/capabilities/built-in-tools/) may be switched off |
| It stopped saving artifacts | The artifact tool group may be switched off |
| It behaves like an older version of itself | A prompt [revision](/docs/build/system-prompt/#revisions-and-rollback) may have been restored |

Changes apply to the *next* run. An agent that ignores an edit you just made is probably mid-run.

## A routine is not running

- **It was paused after repeated failures.** Endue pauses a failing routine and notifies you rather than retrying forever.
- **It ran and refused an action.** Unattended runs cannot send or delete — see [Routines](/docs/automate/routines/#nobody-is-there-to-approve).
- **It ran and you did not notice.** Check the [inbox](/docs/work/inbox/); output usually lands as an [artifact](/docs/capabilities/artifacts/).

<Aside type="tip" title="Reproduce it in a fresh conversation">
  A long thread carries history that changes answers independently of anything you
  configured. If a problem is hard to pin down, try the same request in a new
  conversation before concluding the agent is broken.
</Aside>

## Getting help

- **Search these docs** — press <kbd>Ctrl</kbd>/<kbd>⌘</kbd> + <kbd>K</kbd> from any page.
- **Email us** — [support@endue.ai](mailto:support@endue.ai). Include what you asked for, what the agent did, and which tool call went wrong; that is usually enough to answer without a back-and-forth.

## Related

<CardGrid>
  <LinkCard
    title="Troubleshooting connections"
    href="/docs/connect/troubleshooting/"
    description="Expired tokens, revoked access, the wrong account."
  />
  <LinkCard
    title="Testing and iterating"
    href="/docs/build/iterating/"
    description="Turning a recurring problem into a better agent."
  />
  <LinkCard
    title="Runs"
    href="/docs/work/runs/"
    description="Reading a run, and every reason one stops."
  />
</CardGrid>