For months I'd been losing time to the same slippery pattern. My AI sessions would start strong, then eventually I'd sense something was off. This often traced back to a turn where the agent made a fair assumption, sometimes many, silently leading to slow and steady drift. Before I realized it, there was a noticeable gap between what the agent was doing and my intent at the start.
So I sat down to build something that would catch it for me. A few rounds in I caught the irony - the thing I was building to stop the drift was drifting right before my eyes. I'd assumed it understood me, and it assumed it understood me, but neither of us checked.
This is not surprising if you understand how these tools work. An agent predicts what comes next based on the context it's gathered from your chat, and then builds on that. It is similar to how we interact (ever finish someone's sentence?) but with a couple meaningful differences.
Humans emit unique expressions that help us to contextualize situations. Facial expressions, body movements, voice inflections, conversation speed - these cues enrich our read of a situation and sharpen our sense of someone's true intent. This is why sarcasm doesn't always land over text.
Every living moment grows our baseline. Human experiences form episodic memories that update our contextual awareness and understanding in real time. Agents do not have this feedback loop - they lose it all the moment you end the chat.
A smarter model won't fix this. What makes an answer right or wrong is your intent — and that lives in your head until you say it out loud.
Here are the key takeaways I've gathered from this experience:
It is hard to spot drift up close. Every step can check out but only from a distance can I see how far off-course it went. Now, a fixed point to measure against is non-negotiable — I document explicitly where I am aiming in the first place and what success looks like.
A guess can be a gift or a landmine, and the only difference is the context where it happens. Mid-brainstorm, I want the leaps — surprise me, that's the job. Mid-instruction, that leap is a hidden premise that hardens the wrong thing. So now I say up front which mode I'm in, and have it confirm when it's unsure.
The check pays you back in time. Before I trust an output, I ask: "what did you assume to make this?" The response almost always contains something I never thought to confirm — and catching it there costs seconds, not an afternoon of unwinding it.
This isn't unique to AI. People drift for the very same reasons - things go sideways whenever you hand someone a task and quietly trust that the picture in their head matches the one in yours. The dangerous words are the ones you don't hear, so make sure that you do.
One experiment for you to try - next time it hands you something that matters: ask what it assumed to get there, and what breaks downstream if any of those are wrong. If nothing surprises you, hit reply and show me how you do it.
— D
Voice and direction are mine. Produced with AI.