The Main "Work-Performing AIs" — Agents
Whereas Chat "answers," Agent is the way it actually does the work and brings back the deliverable (see the previous chapter). Here we organize the agents whose names come up at work by their roles. Product names, delivery forms, and pricing change easily, so what to remember is "which suits what."
ChatGPT's Agent Capability (OpenAI)
- Good at: web operation in a virtual browser (research, form entry, applications), file creation, research reports
- Form: integrated into ChatGPT's "agent mode" (the former standalone Operator is consolidated into agent mode); a browser-resident style also exists
- Suits: trip/shopping prep, web-completable applications, cross-site information gathering
- Note: important actions (send/purchase) basically insert a confirmation
Claude's Agent Capability (Anthropic)
- Good at: work involving code execution, routine PC tasks, cautious behavior that asks confirmation
- Form: a "Cowork"-style way of delegating PC work, and developer-oriented Claude Code (multi-file edits, test runs)
- Suits: data formatting/aggregation, routine entry into internal systems, code fixes
- Note: may require OS-permission grants
Manus (General-Purpose Agent)
- Good at: a general type that runs through multi-step, longer tasks
- Form: progress can be checked on the web as it runs
- Suits: time-consuming research, prep for composite tasks
- Note: still weak at fully auto-completing long sequences; check at milestones
OpenAI Codex (For Developers)
- Good at: driving code generation/fixing/PR creation automatically
- Form: a cloud-execution type and a local CLI type
- Suits: bug fixes, test additions, routinizing refactors
- Note: have a human review changes before merging
Judging Which to Choose
| What you want to do | Suited family |
|---|---|
| Web-completable applications/booking/research | ChatGPT's agent capability |
| Routine PC work / data formatting | Claude's agent capability (Cowork-style) |
| Time-consuming composite research | Manus |
| Automating code development | OpenAI Codex (or Claude Code) |
Cautions Common to All
- Supervision assumed: full hands-off is dangerous; a human checks important milestones
- Least privilege: be careful how far you hand over passwords/internal access
- Confidential data: check internal rules before handling it
- Cost: often metered/upper plans; long tasks rack up cost—check the cap. Pricing changes—verify on the official page
- Long multi-step: still weak at end-to-end; delegate in small segments
The First One
For "a light trial," the web-centric ChatGPT agent capability is easy to start; for "solid business use," the confirmation-asking Claude family is reassuring. Engineers, start with Codex / Claude Code. The next article moves to the "first agent" steps to actually run one.
