Two degrees off does not look like a crisis. It looks like a rounding error. A shrug. A "close enough" that lets the meeting end on time. At the start of a journey, two degrees is almost invisible. You and the person next to you are still pointing at roughly the same horizon.
Keep walking — keep shipping — keep generating — and two degrees becomes a different country.
Two degrees off doesn't mean a lot, until you are ten million lines of code down the pipeline.
That sentence is about software. It is also about everything else we are building right now: agents, companies, cultures. Small angular error. Long distance. Expensive arrival.
Why the early angle is the whole game
Downstream is a terrible place to discover you were aimed wrong. By then the pipeline has multiplied the miss: modules depending on modules, prompts depending on prompts, habits depending on habits. Rework is not "fix the last commit." Rework is arguing with a cathedral that was built facing the wrong street.
This is why I care so much about the invisible layer in the demo that is not the product. The demo is day one — still near the origin, where two degrees still looks fine. Production is miles down the line. If your security model, your eval harness, your data ownership, your "what is this tool allowed to do" were two degrees off at the start, you will not notice until the miss is structural.
Alignment is not a vibe you apply at the end. Alignment is the angle you set when the cost of being wrong still fits in a conversation.
The same principle applies to agents
Agents compound. That is the point of them. A slightly wrong system instruction, a slightly soft boundary, a slightly optimistic assumption about what "done" means — and the agent will diligently build you a skyscraper on that foundation. It will not stop to ask whether the foundation was two degrees off. Diligence without alignment is just faster drift.
So the discipline is the same as the codebase: check the angle early. Decompose the work. Write the plan. Execute against it. Revise when reality talks back. When confidence is thin, get a second opinion before the pipeline multiplies the miss. I have been arguing for intelligence over laminated playbooks — thinking that can move when the ground moves — and this is the twin: thinking that still knows which way "forward" was supposed to point.
An agent that is two degrees off will not feel wrong. It will feel productive. That is the trap. Productivity is not proof of aim.
Humans — same principle, greater variance
With humans, the math is harsher. Agents are variable; people are weather systems. Mood, ego, fear, incentive, the story someone needs to be true today — all of it bends the heading. Two people can leave the same meeting with two different norths and both believe they agreed. Six months later you are not "slightly misaligned." You are running two companies under one logo.
That is why bringing a team through an AI transition is not a tooling problem. It is an aiming problem, repeated every week. Truth early, paths for every person, rituals that keep the trendline visible — those are angle checks. Without them, variance compounds into culture, and culture is the hardest pipeline to unwind.
I rebuilt a company from 320 people to 38 on an intelligence layer. Do you think the expensive part was the models? The expensive part was every human heading that drifted two degrees and then kept walking. Agents will amplify whatever aim you give them. Humans invent new aims mid-stride. Both need the early check. Humans need it more often.
Catch the degrees while they are still degrees
- Name north in writing. One sentence for the outcome. If you cannot say it, you cannot notice when you leave it.
- Inspect the first mile, not just the launch party. Architecture reviews, prompt reviews, role reviews — boring on day one, cheap compared to day one thousand.
- Treat "close enough" as a yellow light. Close enough at the origin is how you buy a different destination on installment plans.
- Re-aim on a cadence. Agents get evals. Humans get honest conversations. Both get a second look when confidence is thin.
Two degrees off doesn't mean a lot — until you are ten million lines down the pipeline. The same principle applies to agents. And to humans — where the variance is so much greater. Set the angle while correction is still a conversation. Downstream, correction is an excavation.