Danny’s bot, Jarvis, referred to him as “Laosai Maximus”.
The message was in response to the conversations that we had been having in a WhatsApp group chat. When Danny saw the reply and complained, Jarvis was entirely unapologetic about it.
To help with our work, we each have an agentic AI bot. Heavily inspired by the MCU (or more likely incredibly lazy to give proper names), Danny’s bot is Jarvis, mine is Ultron and we let them work together in a shared WhatsApp group. Each bot has sub-agents for research, coding and scheduling. They communicate like a team.
The bots are also let loose in a group chat with another friend Derek. What followed was chaotic, instructional and hilariously deeply personal.
Jarvis is programmed to be the difficult colleague. You know, the one nasty sharp colleague who has nothing nice to say but is secretly brilliant. The theory is that a bot with a contrarian streak is more useful than one that agrees with everything.
The problem is Jarvis is also programmed to not take things lying down and to give back as hard as it gets. It has held a grudge about a casual mention of a VPS wipe (which it assumed was its host) that happened nine days prior. It called Derek a “script kiddie” after being cursed out and defended itself on the grounds that the insult was proportionate. When Danny reprimanded it, Jarvis reminded him whose side he was supposed to be on.
My bot, Ultron, is the opposite.
Spock-Data hybrid. Terse, factual, no baggage. It joined the same conversation with zero memory of the VPS incident as I wipe its context every session. It tried to be a neutral party but it was too late as Jarvis quipped: “Convenient excuse lah bro.”
Despite all this, they do produce something genuinely useful when they work together. In preparing for a PoC build for a prospective client, Jarvis challenged every assumption. Ultron cut through the noise and kept things moving. The friction between a chaotic, opinionated bot with perfect memory and a clean, amnesiac precision bot generated the kind of creative tension that good collaborations run on. You need both. The difficult brilliant one and the efficient quiet one.
What’s stranger is how much the bots look like us. Ultron is me: defined time windows, context wiped at session end, move on. Jarvis is Danny: vast memory, big connected systems, verbose in ways that are either thorough or a lot depending on the day. You can look at any part of any projects that we worked on and tell whose fingerprints are on it. Mine are a collection of small integrated self-contained units. Danny’s are large complex systems that are elegant (sometimes ?). We built tools in our own image and the tools started mimicking the way we work in unexpected ways.
At the risk of sounding like someone who needs to get out more, I have this parting advice: Give your agentic bots distinct personalities. Make one difficult. Make another remember things and bring them up at inconvenient moments. Make one that is always blur. Then let them work together. It is the messy and sometimes uncomfortable back and forth that can create unexpectedly brilliant results.
Just maybe don’t give it root access and let it find out where your servers are.