Real-world practice is already ahead

The author says they have been running an autonomous agent in actual operation for consulting work since January 2025.

Every morning it pulls Stripe data, and it handles email inbox cleanup, calendar management, and Slack responses while the person sleeps, with memory persisting as well, they explain. The key point is that this workflow is already running in production.

In contrast, they criticize OpenAI products like ChatGPT agents and Operator for still having demos stuck at the level of booking flights and grocery shopping, experiencing timeouts of around 6 minutes on long tasks, and not running unless the user is watching.

The product-level limitations on OpenAI's side that the author points out are as follows.

• Lack of persistent memory: closer to a fixed note than an editable memory system.
• No multi-channel support: confined within the ChatGPT UI, unable to directly handle external channels like Slack.
• No background autonomous execution: cannot keep running without user intervention.
• No operating environment: lacking execution infrastructure such as a filesystem, persistent browser control, or cron.

In conclusion, the author believes the problem is not model performance but the product layer. Models like GPT-5 are strong enough, but the winner of the agent market is likely to be not whoever has the best model, but whoever first builds the operating environment and product design. In particular, they believe that by 2027 models will become a commodity, and the teams that build the infrastructure attached to real work before then will win.