2025 · SEO SaaS, Canada

The agent hung up on people after four seconds

The client wanted prospects called, pitched, and sent a document during the call. Not after. During, while the person is still on the line, because a link that arrives an hour later gets ignored.

That part worked. What did not work was that the agent kept hanging up on people about four seconds in.

What made it hard to see

An outbound calling agent has no staging environment worth the name. You can test against yourself all day and learn almost nothing, because you know to say “hello” clearly into a good microphone after the connect tone. A real prospect picks up in a car, says something that is not quite hello, and there is 300ms of carrier noise on the line first.

The logs were not much help either. They showed a call connecting, a turn completing, and the agent invoking end-call. Every one of those was doing exactly what it was configured to do.

Finding it

I listened to the recordings. That is the whole technique, and it is the thing I would do first again.

On the failing calls, the transcript’s first turn was the single word no. The prospect had said “hello”. Speech-to-text, working on a noisy first packet, had returned “no”, and the agent had a perfectly reasonable rule that a prospect saying no should end the call politely rather than keep talking.

So nothing was broken. Two correct behaviours composed into a wrong one, which is the usual shape of this kind of bug and the reason reading either component in isolation tells you nothing.

The fix

Three rules, and the ordering matters:

The second one is the interesting one, and it is not a speech fix. It is an admission that the transcription will sometimes be wrong, and that the cost of the two errors is wildly asymmetric: hanging up on someone who wanted to talk is unrecoverable, while continuing for one more turn with someone who genuinely said no costs a few seconds and a slightly awkward goodbye.

What else came out of the same sessions

Since I was already listening to real calls:

• Response latency was too high, so I moved the conversational model from gemini-2.0-flash to claude-haiku-4-5. That decision came from calls, not from a benchmark.
• Audio had intermittent glitches, fixed by dropping TTS stability to 0.55.
• The mid-call SMS tool needed its own handling in the close rules, because an agent that hangs up immediately after firing an SMS reads as a bot even when everything technically succeeded.

Compliance was part of the build, not a wrapper

The agent logs verbal consent before sending anything, and handles unsubscribe requests inline for CASL. On the parallel US deployment the same layer covers timezone-aware dialling windows, DNC scrubbing, mid-call opt-out detection, and carrier-spam mitigation so the client’s numbers stay deliverable. I built these rather than buying them, because the rules differ per jurisdiction and the penalty for getting them wrong lands on the client’s phone numbers.

What I would do differently

I would have listened to recordings on day one instead of day three. I went to the logs first because that is the reflex, and the logs were structurally incapable of showing me this: every line in them was correct. The signal only existed in the audio.

I would also now assert on the first-turn transcript in a test harness: feed recorded connect-noise through the pipeline and check that it produces no turn at all. That would have caught this before a real person heard it.

← everything else