The plan changed. Nobody wrote down why.

Making got cheap.

A first draft in a minute. A working branch before lunch. A second one by the afternoon.

The plan didn't get cheaper. It got quieter.

It changes in a thread, a standup, a direct message. The change reaches some of the people, some of the time.

Two plans, then.

The one in the tracker, and the one in everyone's head. They match at the standup and drift by lunch.

This essay is about the gap between the two.

Search interest in AI coding agents rose from 6.69 to 48.77 across four 13-week blocks.

Google Trends, United States, June 2025 to June 2026. Each block's rise was bigger than the last.

Then it cooled 12%.

The summer block fell to 42.85. We say so because the curve does.

And every month, more of the plumbing.

Downloads of the official MCP SDK rose every month from March to August 2026, 4.58 to 6.73 million a day. Installs and CI runs, not people.

Staff used an assistant in about 60% of their work. They could fully hand off only 0 to 20% of it.

132 Anthropic engineers and researchers, August 2025, self-reported. The making moved to the machine. The deciding stayed with people, and the deciding is the plan.

The cost of making went down. The cost of keeping one plan didn't move.

Half of knowledge workers have found another team doing the same project.

Atlassian, 5,000 workers, 2024. That is what a plan looks like when it changed for one team and not the other.

More than 10 hours a week lost to organizational inefficiency, say half of developers.

Atlassian and Wakefield, 3,500 developers and managers, 2025. Their own estimate of their own week.

People made about 70% of the planning decisions and about 20% of the execution decisions.

Roughly 400,000 agent coding sessions, October 2025 to April 2026, interactive sessions only. The plan is still where the humans are. It is also where the record is thinnest.

Priorities changed on Tuesday. You found out on Friday.

Nothing was hidden. The change just never reached the branch you were on, or the agent working it.

A replan that doesn't reach the person inside the old plan is two plans.

Ask why a task was dropped a quarter ago and watch what happens.

Someone scrolls a channel. Someone remembers a meeting. Someone asks the person who has since left. Nobody finds it.

Nearly 20% of the week goes to looking for internal information or tracking down the colleague who knows.

McKinsey Global Institute, 2012, built on IDC's estimates. When messages become a searchable record, that time can fall by as much as 35%.

A quarter later, the branch is merged and the reason is gone.

Git keeps what changed. The tracker keeps that it closed. The why lived in a call nobody recorded.

Every tracker logs what got done.

The skip, the deferral, the reprioritization: those are the decisions that explain the plan, and they usually live in a meeting.

The plan is the decisions. The record usually keeps only the outcomes.

A change to the plan is previewed before it lands.

In Ophanimus Forge, a structural change applies against a preview with a digest. If the graph moved since the preview, the apply is refused and the preview is redone. Nobody's edit lands on a graph they haven't seen.

Work in progress is protected.

A replan cannot silently rewrite a task someone is holding. The holder is part of the check.

And the log is hash-chained.

Each entry carries the hash of the one before it. Last month's reason can't be quietly edited into a different one, and neither can last month's order.

None of that needs a meeting.

The preview is the meeting. Whoever holds the work sees the change, the reason, and who asked for it, from inside the editor they were already in.

When the plan changes, three things travel with it: the change, the reason, and the receipt.

Call it an execution graph: the dependency graph of claimable work that people and their agents execute against.

It knows who holds each task, what is ready, what is blocked, and why the order changed. Forge enforces it and runs none of the work itself.

Once the graph is where the plan changes, the plan becomes measurable.

Not how fast things got made. How long a change took to reach everyone who needed it, and how many it missed.

Three numbers the graph keeps.

Decision latency: from "we should change this" to the change landing in the graph. Reconstruction time: how long a teammate needs to continue a task without rebuilding the plan from chats. Stale-direction count: how many holders were still on the old plan when the new one landed.

108,289 ranked choices. 1,543 decisions with a written reason.

Founder use, not customer data: one person's backlog, March to September 2026. The point is not the size. It is that the reasons exist at all, next to the work they explain.

You can't measure what the meeting ate.