Claude 4 and GPT-5 made agent work a maintenance problem

Claude 4 and GPT-5 pushed agent work into a less glamorous category: maintenance.

That is a sign the technology became operational enough to own. A short answer can disappear into a transcript. A long-running agent accumulates state: files inspected, edits made, commands run, checks skipped, assumptions carried, and approvals granted.

If that state exists only in the conversation, interruption is expensive. The next run reconstructs the work from memory and guesses.

A durable run record can stay plain:

request:
  fix failing markdown validation in blog posts

state:
  editing batch 3 of 8

changed:
  12 files

checks:
  markdown:check passed for changed files

open questions:
  full build not yet run

That is not administrative overhead. It is the difference between a resumable workflow and unmanaged state.

GPT-5 made a related issue visible by making automatic reasoning a default product behavior. Routing the request for the user is convenient, but the choice still needs a budget, a log, and an evaluation. Which tasks got deeper reasoning? Did it improve accepted output? How much latency and cost did it add? Did hard tasks stay on the cheap route?

Automatic does not mean unowned.

Prompts, tool descriptions, policies, and verifier instructions age alongside the model. A model upgrade can turn a previously harmless tool description into a broad permission, or make a verifier miss a new kind of false success. Model upgrades are behavior migrations, not only dependency bumps.

I want an inventory that connects the route to its surrounding assets:

route:
  coding-agent-default

prompts:
  planner-v6
  editor-v4
  verifier-v3

tools:
  file-read
  file-write
  shell-check

evals:
  focused-edit
  no-unrelated-files
  markdown-repair

Agent evals should inspect the path, not only the final prose. Did the agent inspect the right files? Keep scope? Ask at the right boundary? Run the intended checks? Stop when uncertain? Leave a usable handoff?

A plausible completion message can conceal a broken repository.

Approvals need maintenance too. “Allow file edit” is weak when the agent can restructure a directory. The prompt should name the target, show the command or diff, identify the account in use, explain what happens after approval, and describe recovery.

When a model is retired or replaced, treat it like an operational migration: map affected workflows, run local evals, compare cost and latency, inspect changed failure modes, update prompts and tools, review approval copy, and monitor the first production runs.

The model is one dependency in an agent system. The route, prompts, tools, verifier, approvals, run ledger, and cost controls all shape the result. Once agents do real work, keeping that loop healthy becomes part of the product.