Skip to main content
ENES
Bring a workflow →

The agent doesn't respond

Symptom​

In the test chat, after you send a message, you see "Esto está tomando más de lo esperado…" (This is taking longer than expected…), and then "El agente tardó demasiado en responder. Intenta de nuevo." (The agent took too long to respond. Try again.) or "El agente no pudo responder." (The agent could not respond.) with the Reintentar (Retry) button.

Cause​

The agent does three things for each message: it queries Sources, calls tools and calls the model. Any of them can fail or take too long:

  • The model is saturated or the external provider rejected the call (request limit, invalid credential).
  • A tool doesn't respond: Web Search without internet access, an MCP server that's down, a workflow used as a tool that failed.
  • Reasoning effort Alto (High) with a long message exceeds the maximum time.
  • You sent another message while it was responding ("Espera a que el agente termine de responder" (Wait for the agent to finish responding)).

How to confirm it​

  1. Click Reintentar once. If it responds, it was a one-off spike.
  2. Temporarily switch the model to a local InferNode model. If it responds, the problem is the external provider or its credential.
  3. Remove the agent's tools (or lower the Herramientas (Tools) preset to Bajo (Low)) and test. If it responds, a tool is the cause; turn them back on one at a time.
  4. If the agent is a node in a workflow, open the run in Registros (Logs) and read the Traza (Trace): the error event names the node, the tool or the model.

Solution​

  1. External model: check the credential in Administración > Credenciales (Administration > Credentials) (status Conectado (Connected)) and the provider's limits; meanwhile, use a local model.
  2. Tool: fix the tool (S3 or MCP credential, internet access) or remove it if the instruction doesn't use it.
  3. Time: lower the effort to Medio (Medium) or Instantáneo (Instant), and Longitud (Length) to Media (Medium); split very long questions.
  4. If it persists with a local model and no tools, report it with the run identifier (how to report).

Next step​

Last updated on