Agent Plans Need Witnesses

The new trick in AI tools is not only that they can do more work. It is that they are learning to pause and describe the work, with the confidence of a consultant who has discovered Markdown.

GitHub is adding planning agents that explore a codebase with read-only tools and save implementation plans before anything gets changed. The Copilot app is being framed as a place to manage agent-native development, where plans, pull requests, and workflow state become visible objects. Microsoft is talking about "tokenomics" and the question every leader now has to answer: should a human do this, or should an agent? OpenAI is pushing Codex into non-engineering work, with annotations, plugins, sites, and people asking a coding agent to make reports, dashboards, presentations, and lightweight tools.

The theme hiding under all of this is not automation. It is premeditation.

Before the agent acts, it produces a plan. Not a vague intention, but a work object: saved, reviewed, annotated, handed off, implemented, and sometimes used as proof that adult supervision occurred. The plan becomes the little vestibule between desire and execution. A person asks for a thing. The machine proposes a route. Somewhere in the middle, a human is supposed to notice whether it is sane.

That sounds healthy. I like plans. I have written enough software, strategy documents, and mildly overcommitted Tuesday lists to know that a plan can prevent stupid afternoons. A good plan turns invisible assumptions into something people can point at before everybody starts happily improving the wrong thing.

But the plan is also a dangerous shape, because it can make uncertainty look organized.

Anyone who has worked around real organizations knows the genre. The plan has headings, subheadings, a calm little numbered list, stakeholders, dependencies, validation, rollout, and success criteria. It has the emotional temperature of a hotel conference room. By the time a thing looks that composed, disagreeing with it can feel strangely rude, even when the plan is mostly a decorative fence around a guess.

Still, the most important thing about a plan is not whether it is tidy. It is whether it has been witnessed by someone who understands the consequences.

A plan can be technically correct and socially wrong. It can solve the stated request while violating the unstated agreement that keeps a team functional. It can optimize for the shortest path while stepping over the customer support workaround everyone knows is necessary until the contract changes. It can identify all the code files and miss that the person who owns the legacy integration is on leave.

This is where the current planning turn matters. If agents are becoming planners as well as workers, then planning becomes an interface for judgment. The plan is where intent gets translated into action, and translation is where meaning tends to get bent.

Microsoft Research has been arguing for goals as first-class abstractions in human-AI collaboration, which is a wonderfully dry way of saying the quiet part: the tool cannot collaborate well if the shared object of work is mush. Goals, constraints, plans, and checkpoints need forms that both humans and systems can inspect. That is true. The missing piece is that humans also need permission to treat those objects as unfinished, contestable, and a little suspicious.

Without that permission, the planning interface becomes another approval screen with better manners. The user reads a generated plan at 4:42 p.m., sees reasonable nouns in a reasonable order, and clicks proceed because the meeting starts in eight minutes. Later, when the work comes back misaligned, the system has a record. The human approved the plan. There it is, in the transcript, wearing shoes.

This is not a failure of individual diligence. It is a design problem. People are not naturally good at reviewing compressed futures, especially when the future is written by a machine that sounds like it has already been to the debrief. A plan asks the reviewer to imagine consequences that have not happened yet. It requires domain context, memory, taste, energy, and sometimes the social courage to interrupt momentum.

Better planning systems will need witnesses, not just approvers. A witness is not there to bless the document and absorb liability. A witness is there to notice. They ask where the plan got its confidence, what sits outside the context window, what would make the plan wrong, and whether this is the kind of task where speed is how the damage hides.

This changes the product surface. A humane agent plan should show its evidence without making the user dig through a transcript like a tax audit. It should mark what it did not inspect. It should make the riskiest assumptions visible. It should invite comments where humans actually disagree: scope, tradeoffs, naming, customer impact, support load, maintenance burden, reversibility. It should distinguish "I can do this" from "this is a good idea."

It should also make room for human texture. The analyst may know that the clean data source is clean because three people spend Friday afternoons cleaning it. The designer may know that a harmless copy change will make the product sound like it was raised by procurement software. None of that is likely to appear unless the organization has designed ways for lived knowledge to enter the room before the agent starts moving furniture.

There is a hopeful version of this. Plans could make agent work more legible, slower in the right places, easier to teach from, and less dependent on heroic after-the-fact review. They could help teams argue earlier, when the work is still cheap to redirect.

But that only happens if we resist treating the plan as a productivity artifact. The plan is not proof that the machine understood the work. It is an invitation to find out whether the organization understands the work well enough to let the machine touch it.

As agents become more capable, the humane question will keep moving earlier. Not just who approves the output. Not just who audits the action after it lands. Who was present when the work became thinkable? Who had a chance to notice the missing context before execution made it expensive? Who could say, with enough authority and enough time, that the plan was beautifully written and still not quite true?

The future of work may be full of agents that can plan. That does not mean the future needs fewer humans in the room. It may mean we need humans in the room sooner, before the work has hardened into a diff, a deck, or a decision. Witnesses to the fragile moment when a sentence becomes a system.