Claude 4 made long-running agent work a planning problem

Claude 4 made long-running agent work feel less hypothetical.

The important consequence was not that every agent should run for hours. It was that sustained work became plausible enough for supervision to need a design. A capable agent can inspect files, revise a plan, run checks, and keep going. It can also accumulate a wrong assumption while looking productive.

That is a different problem from supervising a chat answer.

A plan for long-running work should create stopping points rather than one giant to-do list:

inspect failure and identify likely files
propose smallest useful fix
apply focused edit
run targeted verification
summarize diff and remaining risk

Each phase gives the user and the runtime somewhere to review evidence or change direction. If the proposed fix suddenly requires six unrelated modules, that is a new planning event. If verification fails, the agent should explain the failure instead of editing forever.

The first phase should often be read-only. A useful inspection result is specific:

found:
  failing markdown rule in three MDX posts

likely cause:
  code fences conflict with repository conventions

proposed edit:
  replace affected fences only

verification:
  pnpm markdown:check changed files

That is the point where permission to edit becomes meaningful.

Supervision should be proportional. Reading files in scope can be automatic. A visible diff makes sense for a named edit. A new directory, dependency change, destructive command, publish action, or external write deserves a stronger pause. The agent should recognize when the run has changed category.

Interruption is normal. The laptop closes, a tool times out, a command hangs, or an approval waits too long. The run needs a handoff before it ends:

current phase:
  targeted verification

completed:
  edited two posts

blocked:
  third post uses mixed MDX component syntax

safe next step:
  inspect the component import before editing

do not do:
  run the full rewrite batch yet

Planning is not there to make the agent sound thoughtful. It shrinks the blast radius by naming the smallest change, likely files, verification, and exclusions:

out of scope:
  no dependency upgrades
  no formatting outside changed files
  no category or tag changes
  no generated templates

Verification should be selected before edits. The check may be a focused test, Markdown validation, screenshot pass, or permission review depending on the work. Choosing it afterward encourages the agent to search for any green result instead of the right evidence.

Long-running should mean resumable and inspectable, not unsupervised. Claude 4 made that distinction more urgent because the model was capable enough to keep going. The plan is what tells the system where it is allowed to stop.