Platform / Limits & troubleshooting

What it can’t do, and what usually goes wrong

The honest list. Knowing these up front saves you designing around a capability that isn’t there — or debugging something that has a one-line answer.

Not supported today

The agent can’t send SMS mid-call

The platform sends SMS, but it is deliberately not a tool the agent can invoke on its own — messaging carries its own consent and registration rules. Have the agent confirm verbally and record what happened; send the message from your own systems afterwards.

Caller ID doesn’t look people up for you

{{ caller_id }} is always available, but turning a number into a customer is your lookup: a Call an API step against your CRM, with the number as input. There is no built-in directory.

Post-call summaries are sampled

Transcripts and the action timeline are captured for every call. Automatic summaries and topic tags are quality-scoring features and run on a sample — don’t build a workflow that assumes every call has one.

No cross-call topic analytics yet

You can read any single call in full. “What did people call about this week?” is not built.

Common gotchas

The agent says a variable as an empty gapThe variable was never captured. Run a Dry run — captures appear inline, so you can see which step failed to save. Check the spelling matches exactly.

It asks the same question twiceIt didn’t hear an answer it could use, or the caller paused and the agent decided their turn was over. Try the Patient caller preset in Voice & turn-taking.

It talks over the callerSame setting, other direction — the agent is treating a pause as the end of a turn too eagerly.

A transfer didn’t connectCheck the destination is E.164 (+14155550100) and that somebody answers it. Transfers can’t work from a browser or Playground caller — test from a real phone. See Transfers.

A proper noun comes out wrongSpeech-to-text guesses at unusual names. Expect it, and confirm important values back to the caller rather than acting on a single hearing.

My agent shows OfflineFor SDK agents, the worker process stopped or lost its connection — an agent is only online while something is running it. See Deploy & hosting.

I see two copies of my agentAn agent is identified by name and environment. A worker booting in a different environment than the one it was registered in creates a second row.

My changes aren’t liveSaving keeps a draft. Publish snapshots a version and Go live points the running agent at it — the dry run uses the last saved draft, which is why it can differ from what callers hear.

🔍

Start with the call itself. Open the call under Calls — the transcript shows what the agent did between turns, with failures in red. Most questions are answered by reading it top to bottom before you go looking at logs.