FIX 01 / 03 — INTENT
Intent
The destination. What success means.
Source — the developer
A talk on AI development
Intent · Reality · Focus — the three fixes that locate AI work.
An agent's output is only as good as what it stands on. Lose any one fix and the work floats — it keeps producing, it just stops being about what you wanted.
34.0489°N · 111.0937°W / FIX 03 / 03
The failure you recognize
It compiles. The tests pass. It's elegant.
And it's wrong.
Your instinct blames the model — not smart enough yet. But the intelligence was never the problem. The agent reasoned fine. It was standing on the wrong thing.
Grounding, defined
Borrow the word from two places. In electronics, ground is the reference everything is measured against — lose it and the numbers stop being about anything. In surveying, three known stations fix any position exactly; two leaves an ambiguity; one, and you're guessing.
FIX 01 / 03 — INTENT
The destination. What success means.
Source — the developer
FIX 02 / 03 — REALITY
The terrain & the map. What is actually true.
Source — the environment
FIX 03 / 03 — FOCUS
The heading. What it's attending to now.
Source — the agent
The claim
Reality scales with money. Intent and Focus scale with discipline. Spend accordingly.
FIX 02 / 03 — Reality
It invents an API that doesn't exist. Wrong signature. A style you abandoned years ago. It breaks a contract it never read. We call this hallucination and blame the model.
"Hallucination is un-paid-for observation."
Reality is observable — the code is right there; the test passes or it doesn't. So it's costly, not hard: effort buys success, you just have to pay. Every bit of reality you don't let the agent observe, it will invent. And this is the corner that's getting solved — use the tools, but don't build your career on it.
FIX 01 / 03 — Intent
Intent lives in your head — partial, mutable, pre-verbal. The agent reads the shadow you wrote down. The gap between what you meant and what you said is where it fails.
"It built the wrong thing, correctly."
Hard, not costly: humans under-specify, assume shared context, change their minds, and don't know what they want until they see what they don't. A better model can draw intent out of you — it can't manufacture what you haven't formed. So author it. Intent is layered: task → goal → constraints. Give only the task and it satisfies the task while violating the goal.
The highest-leverage move in the whole frame: convert intent into reality. A test or a type is intent pre-compressed into an observable form — a good test moves a slice of what you want from the side money can't buy onto the side it can. When you re-explain the same intent a third time, stop explaining and encode it.
FIX 03 / 03 — Focus
It can fail when everything else is perfect. Rabbit holes. Forgotten constraints. Over-engineering. Drive-by changes. Quitting early — or never stopping.
"Focus gets worse as you spend more."
More context is more room to get lost; more autonomy is more chances to drift. The resources that buy reality grounding actively erode focus — they pull opposite ways. It's the failure mode of the near future, because we're scaling exactly the dials that strain it. Focus is the agent's discipline, but you design the conditions.
The diagnostic
When an agent fails, don't ask "is the model good enough?" Ask which fix came loose — each failure has a different remedy.
As models improve, reality grounding approaches free. The enduring craft — the part still hard when the models are ten times better — is the two corners compute can't buy.
Intent you author. Reality you observe. Focus you defend.