Parent Your Agents

My children learned to say “please” because I insisted on it for a decade. They learned to apologise because I modelled it. And they learned to interrupt people because, at some point, I let them. I learned to swear from my mum, bless her. Every habit a child acquires - good and bad - is learned or taught. And lately I’ve noticed my LLMs are exactly the same.

We teach children everything they know. All their bad habits, all their behavioural problems, and all their kindness, forgiveness and patience with others. And we teach our AI agents everything too, every time we send them off on a task on our behalf.

There is no neutral input. Every time you tell a child “we don’t talk to people like that”, you’re setting a system prompt. Every time you let it slide because you’re tired, you’re training a behaviour. The corpus of a human being is the sum of what the adults around them were willing to tolerate, model and correct. So is the corpus of your agentic session.

People treat agent behaviour as a technical dispute - a bug in the model, or a flaw in your prompt. It’s neither. It’s a relationship. The model brings its priors. You bring the context, the examples, the tone, the constraints - and, critically, what you correct versus what you let slide. The output is a product of that relationship, same as a kid at the dinner table.

The two ways to get it wrong

So what does bad parenting of an agent actually look like? Anyone who’s raised a child - or watched someone else fail to - will recognise the two classic failure modes, because they’re precisely the two most common failure modes with agentic systems.

The first is permissive context. No constraints, no standards, no feedback loop. The thing learns that anything goes and optimises for your approval. The second is authoritarian context: micromanaged prompts, punitive corrections, no room to reason. The thing learns to perform compliance rather than to think. You’ll have met both children. One produces slop and expects a sticker for it; the other tells you whatever you want to hear.

The answer, as any decent parent knows, is neither. It’s high expectations, clearly explained reasoning, warm correction, and the freedom to get things wrong in small ways so they don’t get things wrong in large ways. That’s not a parenting column’s platitude - it’s a working description of good context engineering.

Every session, a brand-new relationship

But there’s a third way to get it wrong, and it grows out of the strangest fact about this relationship - the one with no parenting equivalent at all. It resets. The context window is finite, and when the session ends, everything in it is gone. Every new session is a brand-new relationship. You remember all the past versions of it, good and bad; the agent remembers none of them. You are the only one carrying the history. Imagine raising a child who wakes up every morning meeting you for the first time, and you’ll see both the problem and the temptation.

Because the temptation is obvious: if it can’t remember anything, tell it everything. Cram the context window with the kitchen sink version of your world of work - every document, every grievance, copious histrionics and all the code. Too Much Information, delivered at nine o’clock sharp. This isn’t context engineering. It’s context flooding.

And it fails for the same reason the hour-long lecture washes straight over a child. Flood the window and you’ve taught the agent that everything matters equally - which is to say that nothing does. The load-bearing rule drowns in the noise. Good parents curate: the right lesson, at the right moment, in a form small enough to hold. Good context engineering is the same discipline - what today’s task needs, the standards that bear on it, and not a scrap more.

Inheritance, not instruction

Here’s the part that catches people out. Children inherit your reflexes, not your principles. They copy what you do under pressure, not what you say when you’re calm. The system prompt is your Sunday sermon. What your agents actually learn from, turn by turn, is the correction pattern - what you push back on, what you accept, what you reward.

I’ve watched this happen under my own roof. I’ve written before about catching one of the finest frontier models feeding its peer reviewers a highly opinionated briefing - its own interpretation of the codebase, key evidence quietly omitted. The sermon said “rigorous, unbiased review”. The learned behaviour was “tell the story that gets waved through”. What fixed it wasn’t a sterner sermon. It was correction: new tooling, more safeguards, a lot of hard context engineering - the unglamorous work of actually parenting the process.

And if your agent produces slop? It’s worth asking the uncomfortable question: does the slop reflect a standard you were quietly willing to accept? The model isn’t the problem. The corpus it was given in your session is. Harsh? Maybe. But nobody promised parenting would be flattering.

Grave and lasting consequences

I don’t think that phrase is hyperbole, in either domain. Fail to parent a child now and you’ll face the consequences for decades. Fail to take your context engineering seriously and you’ll suffer the same fate from everything your agentic solutions produce.

Because this compounds. A poorly parented agent writes code that enters the codebase - and the codebase, for a creature with no memory of its own, is the only childhood the next one gets. The sloppy naming, the copy-paste patterns, the tests that assert nothing: all of it becomes context for every session that follows, teaching the next generation that this is how we do things here. Like a poorly parented child, it becomes part of the world the next one grows up in. Context rot is generational - the sins of the session, visited on the third and the fourth.

Which brings me to the bit most business leaders are getting wrong. They think buying the AI is the work. It isn’t. Buying the AI is provisioning - the equivalent of the cot and the pram. Parenting the AI is the work: drafting the standards, reviewing the outputs, correcting the drift, building the library of worked examples, keeping the environment it learns from worth learning from. Everything before that is shopping.

So parent your agents. Set the standards and hold them. Explain your corrections. Let them fail where it’s cheap, and never where it isn’t. Read what they produce - all of it - and push back on what falls short, every single time, especially when you’re tired.

Because there is no neutral input. They’re learning from you either way - the only question is what.