What agent mode is
ChatGPT’s agent mode lets a conversation act: the model gets a working environment, takes steps toward what you asked, shows you what it is doing, and hands the result back into the chat. You can watch, interrupt, and redirect. It is a genuine agent by the plain definition — a job, tools, a stopping condition — delivered inside a chat session.
This article is not a takedown. OpenAI built a good on-ramp, and if agent mode is where you first felt the shift from “AI that answers” to “AI that does,” that instinct is correct. The question is what happens when you act on it seriously.
What the chat-session shape does well
- One-off jobs with you present. Book-shaped research, a comparison you need this hour, a task you can describe but not be bothered to do — agent mode shines when you would have stayed at the keyboard anyway.
- Zero setup. No roles, no configuration. You type, it works. For a first contact with agent behavior there is no faster door.
- Live supervision. Watching the agent work is the fastest way to build intuition for what these systems can and cannot do — the same reason the operator’s guide says to run your first missions while you watch.
Where the shape runs out
The limits below are structural — properties of work-inside-a-chat, not of the model. A better model next year inherits all of them.
One session, one job
A chat session holds one thread of work with one worker. The moment you have three delegated jobs, you have three tabs — and you are the only place their status comes together. There is no roster, no “what is running right now,” no shared picture. You became the coordination layer without anyone asking you.
The work ends when the chat does
Agent-mode work is session-bound: it happens while the conversation is open and lands in a transcript. Recurring work — the Monday competitor watch, the standing security sweep — has no home in a shape you must personally reopen. A transcript is also a weak record: scrolling a chat to find out what was done, when, and from which sources is not a system of record, it is archaeology.
Nobody else can see it
A chat is private by construction. That is right for exploration and wrong for operations: the moment delegated work matters to more than one person, “what did the AI do this week” needs an answer that is not forwarding transcripts. Workspaces, a shared map, and per-run receipts exist so that the answer is a surface, not a retelling.
The worker has no job description
Each session starts as a blank generalist. There is no durable role that says what this worker may touch and what done looks like for its kind of work, no action list you can read before delegating, no separation between workers that read and workers that change things — the properties that make delegation reviewable in real agent rosters.
What a managed workforce changes
A managed workforce keeps the part agent mode got right — the model does the work — and moves it out of the chat session into a structure:
- A roster instead of tabs. Named roles with one-line job descriptions and enumerated actions, hired once, on a map you can look at. Who exists and what each may do is a fact, not a memory.
- Missions instead of messages. Work is an assignment with a done-state that runs to done or fails visibly — not a conversation that trails off.
- Schedules instead of reopened chats. Recurring work runs on its cadence and lands in review. You review outputs; you do not re-initiate them.
- Receipts instead of transcripts. Every run leaves a record — what was asked, what was done, what came back — kept the same way for failures as for successes, in a log that is not a scrollback.
The trade is real: a workforce asks for setup a chat never does. You connect provider keys, hire roles, write missions. That cost buys you the thing the chat shape cannot offer at any price — many jobs, visible in one place, with a history you can check. Autonomy you can watch is the longer argument.
The builder road, briefly
OpenAI also ships developer-facing tools for constructing agents in code and wiring them into your own software. That is a third path, and the same honest split applies to it as to Anthropic’s SDK: build when the agent is part of your product or needs tools only your codebase can provide; configure when the goal is getting real work done this week without writing the loop yourself. The long version of that decision — written for Claude but true across providers — is in Claude agents without writing code.
A concrete week, both ways
The difference is easiest to feel as a week. In the chat shape: Monday you open a session and run the competitor check; Wednesday you remember to rerun it, in a new session that knows nothing of Monday’s; Friday you scroll two transcripts trying to recall which claims came from where. Every job happened — and every job cost you the remembering, the reopening, and the reconciling.
In the workforce shape: the competitor watch is a scheduled mission on a research role. It ran Monday and Thursday on its own; both receipts sit in review with sources attached; the map showed the runs happening while you did something else. The model did the same quality of thinking in both weeks. The structure around it is what changed — and the structure is the part you cannot get by opening the chat more diligently.
When agent mode is still the right call
Plainly: if your delegated work is occasional — a few one-off jobs a week, each fine to supervise — agent mode inside the subscription you already have is the right tool, and adding an orchestration layer would be ceremony. The workforce shape starts paying rent when the jobs become standing — recurring, plural, worth a record — or when you notice you are the status meeting between five chat tabs.
Using both without contradiction
This is not a subscription war. Ahel runs on provider keys you already hold — OpenAI’s among them — so the model access you pay OpenAI for backs both your chats and your workforce. Exploratory, while-you-watch jobs stay in the chat; standing jobs with done-states move onto the roster; both draw on the same account, with no markup added on your model usage (pricing explains what Ahel does charge for). Ahel is not affiliated with OpenAI — their models, your subscription, our coordination layer.