#6277: How to coach an agent
I titled the post as I did because lately I’ve been feeling a bit Toothless. A cursory glance at my LinkedIn account will reveal my general bullishness towards agentic AI in the programming/productivity space—and certainly I am less enthusiastic about gen-AI in other more creative contexts—but of course a bull in a china shop doesn’t stop to think about the broken porcelain. Other people have to do that. So I am both that person and the other person right now.
My career hasn’t exclusively consisted of programming, or being a software at the junior, senior, staff level, or various kind of leads across the stack. It forms a large part of my experience but a lot of it is heavily informed by what other experienced people have taught me about agile, management, and how to coach or mentor others.
Are you expecting me to write at length about how to be a good coach, or a mentor? I hope not, because you’ll be sorely disappointed. About as disappointed as I’ve felt, maybe.
The problem is the conversational interface that comes with every LLM since the announcement of ChatGPT. AI in the year of our lord 2026 is not an API call behind a SaaS vendor, and it’s not a workflow orchestration toolkit like Zapier or IFTTT. It’s a thing you can talk to and it will talk back to you either in text or with a synthesised voice that is lightyears ahead of what Stephen Hawkins chose for himself. You can tell it to use a certain style of communication if you really care about it, and it will try to do it. The AI voices I can choose from in Siri and Google all have softly spoken Northern accents (but with none of the dialect, natch).
Effective coaching, mentoring, and leadership for that matter, requires strong interpersonal skills and much of this depends on how well you get on with the person, or people, you’re working with. I’ve had a few managers I’ve totally hit it off with and we’ve worked amazingly together, but there are far more times where it doesn’t gel and the job becomes more difficult through a lack of seeing eye-to-eye, or even people pleasing or saying what you want to hear.
I don’t believe there’s a way to un-murk those waters because human interaction is intrinsically fuzzy and a lot factors—emotional, psyhological, external—can change the nature of those interactions from one moment to the next.
Except an LLM can pull it off because despite being non-deterministic it can be made predictable. It can be made to be a friend, or a challenger, or any other personality you can think of.
Pulling this back into the real world for a minute, I’ve felt the part of my job that involves leadership, coaching, and mentoring, has become intensely more difficult than it used to be. My competition is an AI agent who probably has a lot more answers than I do, even without context, and where once I would create availability for people to talk to me to assist on various topics, those moments are much more difficult to capture these days.
The good news though is that time is valuable and an LLM or AI Agent can do a lot of the trivial legwork without requiring any intervention. This is a pure boon for productivity in the same way Stack Overflow was when it launched back in 2008. It saves time on both ends so you can focus on the deeper, more serious stuff.
At the same time, I’m worried that the chat-based interactions with LLMs are so easy-going and appeasing that they take the place of the people who are invested in sharing or passing along their experience, or navigating the choppy waters of a difficult conversation.
To relate a personal experience, I have a hard time being direct to people who appreciate a direct style, and under the hood I’ve got some serious people pleasing, try hard tendencies. My inclination is to avoid that kind of interaction because it is personally uncomfortable to me even though it benefits the person receiving it. My running assumption is that everybody understands the world the way I do so when I talk about it on my terms, everybody knows what the deal is, as if I can empathise with absolutely everything except the wavelength I operate on. This is obviously untrue and to hold to it would only serve to frustrate a conversation.
And what can any of us do about that? I think you have to avoid being sucked into the matrix, where the machines can show you everything you want to see, make you comfortable, and instead prioritise those visceral experiences not just of success, but of failure, embarrassment, shame, and otherwise simply falling short or fucking up as you carve your way towards a mutual understanding with someone. If not friendship or more. All the good connections survive that growth.