Troubleshooting

Nothing happened at all

No session was created, so the trigger never fired.

• Channel triggers — is the channel connected and its credential valid?
• Interval triggers — is the module in production mode, and is the interval what you think?
• API triggers — is your caller reaching the endpoint at all?

The workflow ran but the customer saw nothing

Almost always a trigger and response mismatch. A WhatsApp trigger with a Chat response runs perfectly and answers into the generated UI, where no customer is looking.

Fix — Give every trigger its matching response node. See Response nodes.

A compiled log showing each node with its resolved template values, one field visibly empty

Screenshot to be added

Compilation exception in the logs

Two usual causes:

• A {{CREDENTIAL ...}} expression written by hand. Pick the credential from the node's dropdown instead — Flux fills it in correctly.
• A malformed expression — a missing brace, or a keyword that is not one of the six.

An AI node answered something nobody asked

Its input resolved to empty. A POUT path that does not match the previous node's actual output produces nothing, and the model answers the empty prompt anyway.

Fix — Open the compiled log and look at what the node received. Then drag the key from Response JSON Keys rather than typing the path.

The agent ignores its tools

The model chooses tools from their descriptions, so the descriptions are the problem.

• Use DESC on every parameter, and say what the value is and what format it takes.
• Two tools with overlapping descriptions confuse the choice — sharpen or merge them.
• Too many tools makes the decision harder. Split the workflow, or route with a Decision node first.

The agent invents answers

• Attach a Knowledge Base, and say in the system prompt to answer only from it.
• Tell it explicitly what to do when it does not know — otherwise it will guess.
• Give it a transfer tool so "I do not know" has somewhere to go.

It forgets what was said a moment ago

No memory attachment. Attach GPT Memory or Gemini Memory — matching the model — so the conversation carries through the session.

A node did not run

Usually a Decision branch that was not taken. Only the taken path executes, so nodes on the other branch are absent from the log rather than failed.

A tool runs every time when it should not

It is wired into the AI node's input instead of its tool connector. On the input it is part of the chain; on the tool connector the model decides.

Cost is higher than expected

• Check the interval triggers first. A short interval left on after testing is the most common cause.
• Long conversations cost more per message — memory sends the whole history each time.
• Every tool is described on every call. Remove the ones you do not use.

See Usage and analytics.

The agent stopped answering

If it runs on Flux's built-in keys, check the credits wallet. An empty wallet fails at the AI node, which looks like the agent going quiet rather than like an error.

Still stuck?

Contact support with the project, the module, and the session from the chat log. The compiled log for that session is the single most useful thing to include.