CodexWorkA familiar way to use AI at work is one exchange: paste text, get an answer, judge it by reading it. Even then, an answer can hide errors. When the work moves into files, apps and repeated runs, there are more places to inspect.
The recent product notes describe something different. The work spans several files, several apps and several days. Its failures are quieter. The agent can’t see the folder you meant. It forgets what it decided last week. It drafts a message nobody sends. It reports “done” while the deliverable sits somewhere you never look.
We think a better model doesn’t remove those problems. A smarter model with no access to the input is still blind. A recurring task without a reliable record of earlier runs risks repeating or missing work. These are properties of the setup, not of the intelligence.
What the vendors say changed
All of this is documentation. None of it is evidence of results in your account.
- OpenAI weekly update. The September 28–October 2 update describes Dots that take on ongoing responsibilities. They work through connected apps, their own computer, and delegated Work or Codex tasks. Access rolls out by plan and region. The same update says Space supports Pages and files. A product note like this doesn’t tell you whether your account has any of it connected.
- Claude release notes. The Claude release notes list Cowork capabilities in any conversation on September 16, along with editable designs, decks and documents. On September 25 they list plugin submission and usage analytics. The notes suggest that the older split between a “chat mode” and a “work mode” was not stable. Don’t build a team habit around that split.
- Cowork scheduling. The current Cowork scheduling page says recurring tasks run remotely. Legacy Pro and Max schedules created before October 6 stay local. Tasks that need local files or apps still need a connected desktop. So don’t assume closing a laptop is always safe. Read the line for your own task.
- Grok Bot launch. The Grok Bot launch on August 11 describes a beta with a persistent cloud computer, cross-app work, learned routines and coordinated bots. It’s a launch post with vendor demos. It doesn’t show independent impact.
- Muse Small Business. Muse Small Business (September 29, updated October 8) describes business connectors and skills for US and Canada. The company states that approval comes before anything is published, sent or spent. That is a statement of design intent. You still need to test it in your own setup.
We aren’t claiming mass adoption. These are announcements. What they share is the shift in shape, and that shift is what matters here.
Three kinds of work, three different specifications
Not every job needs the same handoff. This table shows what a specification must add as the work stretches out.
Scroll sideways to compare columns.
| Work | Example | Specify | Inspect for |
|---|---|---|---|
| One task | Clean this month’s expense export | Input file, output format, exception rules | Assumptions hidden in the data |
| Ongoing responsibility | Watch three product pages weekly | Cadence, earlier state, finish rule, recipient | Stale state or a missing report |
| Shared tool | A hosted page the team opens | Access, storage, allowed actions, owner | Host differences and unclear ownership |
A task ends when the file arrives in the agreed location and passes its usability checks. A responsibility ends only when someone decides it should. That second sentence is the one to reread. If a job has no end, the specification needs a place to keep state and a rule for stopping.
The four things that now carry the weight
- Input access. Connecting an account is not the same as being able to read a source. “Connected” tells you a login exists. It doesn’t tell you the agent can open the shared drive subfolder you meant. Name the exact location and test one read.
- Durable state. Say where last week’s result lives, in what form, and who updates it. Product memory may help, but a dated record you can inspect and pass to another receiver makes continuation easier to verify.
- Approval. List the actions the agent may take alone and the actions that need a human. Reading and drafting are usually safe. Publishing, sending and spending are different. Vendors differ on how they enforce this, so write your own rule too.
- Actual delivery. “The job started” is not “the output arrived.” Define the delivery location and check it.
The tooling layer: don’t conflate it
Three things get lumped together, and they do different jobs.
- MCP Apps let tools present interactive interfaces. What you get depends on whether your host supports them.
- Sites in ChatGPT covers creating hosted apps with durable storage.
- WebMCP exposes tools at the level of a page or session.
Choose the surface by the job. A shared tool can be useful when colleagues need to enter data and return to saved results. An interactive tool interface can help when the action belongs inside an agent conversation. A document or downloadable folder can be enough for a one-off handoff. Specify the reader, action and storage need before choosing the protocol.
The Agent Skills specification points the same way. It describes instruction packages with optional resources and scripts. Host support varies. Treat a skill as a portable instruction, not a guarantee of behavior.
Where a curated pack earns its place
A capable general agent can write its own plan and template. If you ask it, it will. So a prepared pack has no special power, and it works no better because of clever wording.
A pack is worth using in only two cases:
- It reduces setup and recovery work you would otherwise do, such as having sample inputs, an output example with its evidence label and repair notes ready when a run goes wrong.
- It gives you inspectable examples, so you can compare the agent’s output to something concrete.
That is the intended role of the CodexWork task packs. Output examples are authored examples unless a page links a recorded run; inclusion in a pack is not a test result. Each has an instruction, an input template, synthetic samples, an output example, and checks with repair notes. If a pack doesn’t save you effort, write your own. It’s your call.
A first move
Take one recurring job you do. Write one sentence for each of the four items above. If you can’t, that’s the gap. Chapter 2 shows what inspection looks like on a real file. Chapter 3 shows how to hand over a job that repeats.
Scope and sources
Limitations. This chapter relies on vendor documentation checked October 13, 2026. Features, plans and regions change. We tested none of these products’ claims, and we know nothing about your account’s access. Our opinions on specification are judgments, not measured findings.
Next task: Run the spreadsheet cleanup pack on its synthetic sample before using your own data. Or continue to the recorded exceptions.
Source note. Linked provider pages were checked October 13, 2026. They document product intentions and access conditions, not comparative results. The specification table and selection guidance are CodexWork editorial judgment.