AI agents still need babysitting

The honest take: fully babysitter-free agents are still the exception, not the default. The systems that work in production usually look more like disciplined workflows than one magical do-everything model.

• –OpenAI’s agent docs explicitly point to context compaction and narrow scopes as the fix for long-running tasks: https://openai.com/index/equip-responses-api-computer-environment/
• –OpenAI’s internal data agent relies on retrieval, human annotations, and layered context rather than a single monolithic loop: https://openai.com/index/inside-our-in-house-data-agent/
• –The winning pattern is planner/executor separation, explicit checkpoints, and hard stop conditions, not “let it run and hope.”
• –If an agent keeps drifting, the problem is usually not just model quality; it’s weak task decomposition, missing memory, and poor tool boundaries.
• –The next leap is less about raw autonomy and more about reliability engineering: evals, compaction, retrieval, and failure recovery.