Claim
Work/Cowork agents return a finished product for you to review; Code/Codex agents show you the files changed, commands run, and tests performed, making the working process itself the reviewable artifact.
Mechanism
Work and Cowork run on company servers and optimize for a finished end product: a presentation, analysis, or organized file collection. Code and Codex run locally on your machine and expose every intermediate step, producing a commit history, test log, and command transcript alongside the deliverable. Local execution also enables multi-file projects across longer sessions. The choice determines what you are able to audit: you judge the result in Work mode; you audit the process in Code mode.
Conditions
Holds when:
- You need a reviewable output with no interest in the intermediate steps: Work/Cowork.
- The task involves code, multi-file changes, or test execution where the process itself must pass inspection: Code/Codex.
- Compliance or audit requirements demand a record of what changed, not just the final state: Code/Codex.
Fails when:
- Code/Codex's local execution environment lacks the tools or dependencies the task requires.
- You have Code mode's process transparency but do not review the intermediate steps, making the visibility pointless.
Evidence
Work/Cowork: "you ask for a presentation, analysis, or organized collection of files, and the agent returns something for you to review."
Code/Codex: "expose the work itself: the files being changed, commands being run, tests being performed, and a detailed record of the changes." They "run on your computer rather than company servers," enabling "more complex multi-file projects over longer periods."
Signals
- Work mode: you review and accept or reject the finished artifact.
- Code mode: you read the commit diff and test output before merging or deploying.
- You can say in one sentence why you chose the modality for each delegated task type.
Counter-evidence
The Work/Code modality split is a current-generation product distinction, not a permanent architectural one. Future agentic systems may merge both postures. The meaningful decision may eventually reduce to scope of output and depth of audit, not which product you open.
Cross-references
- Agents differ from chatbots because they give the AI a computer to use: the broader frame for why agents run differently from chatbots and why the process vs. output question arises.