Five Watchdogs

The third candidate for the role of Asya's brain was Gemini. The pilot lasted twelve days — and they were the most eventful twelve days in the project's history. By the end, just to keep the engine from dying mid-call, five watchdog timers stood guard around it simultaneously. Spoiler: the watchdogs didn't save it.

On paper everything looked serious. Behind the engine — the company with the largest data and compute on the planet; the price — five cents a minute, exactly our budget. We wired it into the telephony and began teaching it what its predecessors already knew: recognize the caller, check the plan, transfer to a human, say goodbye.

The first surprise: the engine silently discarded part of our instructions. There is a service channel through which an agent receives the context of a specific call — the caller's number, the route, the technical plumbing. Gemini simply didn't read that channel. No error, no warning — the data vanished into the void, and we spent days guessing why the agent didn't know what it had explicitly been told.

The second surprise came from the realm of mysticism. A session would die right after the greeting — total silence to the end of the call — if the description of one of its tools was written in Cyrillic across several lines. Rewrite the same thing in Latin script on a single line — it works. We never did find an explanation; we simply accepted it, like the weather.

The third surprise would have cost us customers, had it reached the production line. Our instructions contained a training example: "a call from a number, for example, 067X...". Gemini did not understand it was an example — and began genuinely looking up subscribers by the training number from the instruction. And once, failing to learn a person's name, it wrote the phrase "Customer declined to give their name" into the name field of a real work order — and submitted it. Formally — impeccable. In substance — a work order from a ghost.

Then came the freezes. The engine froze mid-sentence, froze after calling a tool, froze without waiting for the caller's first word. For every kind of freeze we assigned a separate watchdog — a timer that nudges the agent in the ribs if it stays silent suspiciously long. By the end of the pilot there were five watchdogs. Five nannies around one voice — just so it would survive to the end of a conversation. And it never did learn to say goodbye out loud: it hung up in silence, a limitation we never managed to break through.

A separate discovery was the memory ceiling: the advertised context size turned out, in practice, to be four times smaller — without special configuration the session simply crashed with an error once the conversation ran long. We worked around that too. We worked around everything — and that was precisely the trap.

At some point we stepped back and looked at the construction from the outside: five watchdogs, detours instead of the service channel, instructions in single-line Latin script, examples scrubbed from the prompt so they wouldn't be taken for truth. Each crutch on its own — a trifle. Together — a structure held together with duct tape and hope that collapses at a sneeze. And all of it for the sake of an engine that did nothing its competitors didn't do for the same money — without the five nannies.

We closed it without drama, in a single decision together with the next hero of our series. That conversation will be short — because the romance, too, was the shortest in our history: we passed the verdict in under two days. Grok — in the next dispatch.