Plan in Pencil

I can no longer plan anything past the next month or two, and I've stopped pretending otherwise. There's a tidy irony in why. METR measures how long a task AI agents can complete on their own, and that horizon now doubles roughly every four months — it used to be seven. The machines' horizons are stretching at the exact moment ours are collapsing. They can finally see further; we suddenly can't.

Here's the mechanism. A plan is a stack of assumptions — what things cost, what's possible, who you need — and every assumption has a shelf life. Planning worked for decades because assumptions outlived the plan. What broke isn't ambition or discipline; it's that the half-life of an engineering assumption dropped below the length of the planning cycle. A six-month roadmap written today encodes today's model capabilities. By delivery, three model generations have shipped, the build-versus-buy math has flipped twice, and a two-person startup has shipped your Q4 as a feature. Eisenhower's line — plans are worthless, planning is everything — used to be a paradox. Now it's operating guidance.

The same forecasting error runs in the other direction, and it wears a suit. Executives are pricing in three to five times the output from shrinking teams — Salesforce hired zero new engineers last fiscal year, ServiceNow won't backfill attrition — which is drawing a straight line through the same exponential the dead roadmap died on. The gains are real but they arrive jagged: huge on greenfield, modest on legacy, negative during the learning dip. The quota arrives smooth and immediate. The tool doubles you, the quota triples you, and the difference gets extracted as anxiety unless you refuse the panic and bank the surplus instead.

So what do you do when you can't see past sixty days? You separate what's durable from what's disposable. Direction is durable — your users, your domain, your taste, the problems that will still matter in five years. Routes are disposable. Fix the direction in ink and re-derive the route monthly, without guilt. Shrink your commitments, not your ambition: prefer decisions you can unwind, rent what you used to build, and hold options instead of positions — run five cheap probes, kill four, double the winner.

And exploit the strange gift hiding inside the chaos: re-planning is nearly free now. Long plans persisted because planning was expensive, so we amortized it over quarters. But a plan is tokens now. If your context is written down — the decisions, the constraints, the why — a fresh plan that accounts for this week's world is a prompt away. Agility stops being a ceremony you perform and becomes a property of your context hygiene.

Don't mourn the roadmap. It was always a comfort object; the confidence was the product, not the accuracy. Write plans you're happy to erase. Treat estimates as expiration dates. Ink the direction; pencil the plan — and sharpen the pencil often.