ChatGPT Work promises hours-long autonomy. Here's what breaks first.

OpenAI's pitch for ChatGPT Work is that it finishes whole jobs, not answers. That's the exact claim that falls apart in production, because the failure mode of long-horizon agents isn't reasoning — it's error accumulation over dozens of tool calls. If you've run any multi-step agent for more than 20 minutes, you already know where this goes.

OpenAI debuted ChatGPT Work as an autonomous agent built to execute complex, multi-hour workplace tasks, landing alongside the GPT-5.6 rollout, per Bloomberg's report. The demo framing is "hand it a job, walk away." The engineering reality is more complicated.

What actually breaks in an hours-long agent?

State drift. An agent running for two hours makes hundreds of decisions, and each one is conditioned on a context window that's slowly filling with its own prior output. By step 40 the model is mostly reasoning over its own summaries of summaries, and small misreads early compound into confident nonsense late.

Tool-call brittleness is the other killer. A real workflow — pull data from a CRM, reconcile it against a spreadsheet, draft an email, file a ticket — touches five systems that each fail in their own way. An auth token expires. A pagination cursor returns partial data. The agent doesn't know it got partial data, so it proceeds as if the job is done. That's not a hallucination bug, that's a systems-integration bug wearing an LLM costume.

The subtle one: no clean rollback. When a human workflow goes wrong at step 30, you back up two steps. An agent that already sent the email and filed the ticket can't un-send them. "Finishes whole jobs" means it also finishes wrong jobs, at machine speed, before anyone checks.

Does the autonomy claim hold up?

Depends entirely on what "job" means. Bounded, idempotent, single-system tasks — reformat these 200 records, triage this inbox by rule — are genuinely close to solved. The value there is real and I'd deploy it.

The "multi-hour, multi-system" framing is where I'm skeptical. Not because the model can't reason, but because reliability at that horizon is a math problem: 95% per-step success across 40 steps is a 13% chance the whole run completes clean. You need per-step reliability in the high nines before end-to-end autonomy beats a human with a checklist. Nobody has shown that number yet, and OpenAI's launch didn't either.

Should you use it?

Use it for narrow, reversible, well-instrumented tasks where a bad run is cheap to catch and cheaper to redo. Pilot it there, log every tool call, and build a diff/approval gate before any write action hits a real system.

Don't hand it a genuinely open-ended multi-hour workflow and walk away, no matter what the launch page says. The verdict: promising infrastructure for supervised automation, oversold as autonomy. I'd run it in the loop, not out of it — and I'd hold judgment on GPT-5.6 until the failure-rate numbers show up in someone's postmortem instead of a keynote.

FAQ

Is ChatGPT Work different from the existing ChatGPT agent mode?

The framing is longer horizon and full task completion rather than assisted steps. In practice the hard part is the same — reliably chaining tool calls across systems — so treat it as an evolution of agent mode, not a new category, until benchmarks say otherwise.

What's the biggest risk running it on real workflows?

Irreversible actions. An agent that sends emails, files tickets, or writes to a database can complete a wrong run before anyone reviews it. Gate every write behind an approval step or a dry-run mode.

How long can it actually run reliably?

That's the unanswered question. Reliability degrades with step count because errors compound, so the honest answer is "as long as your per-step success rate and rollback tooling allow" — which is usually shorter than the marketing horizon.