Skip to main content

Where AI Agents Still Need Adult Supervision

OpenAI says its internal coding agents now finish many research tasks alone, but the uneven results are the real lesson for businesses putting agents into live workflows.

Metomorph Editorial 4 min read
in X @

If you're rolling out AI agents, the hard question isn't whether they can do useful work. It's which work you can safely let them finish without someone checking every step.

That's where many teams get stuck. A tool that drafts, monitors and summarizes well can still break the moment it has to build process, handle edge cases or make a messy operational judgment.

OpenAI's own workflow data shows a jagged pattern

The most useful part of OpenAI's new internal report isn't the headline that agents now finish more work alone. It's the admission that the gains are uneven: strong on monitoring, explanation and review, weak on building the systems and environments around the work.

As PPC Land notes, OpenAI is measuring its own researchers with its own tools, so this isn't an external audit. But it's still a rare look at where a leading AI lab trusts agents, where it doesn't, and how fast usage rises once the tasks are well chosen.

  • More tasks finished solo. OpenAI reported that its internal coding agents completed 46% of research tasks estimated at two to four hours with zero human intervention in October 2026, up from 23% in March.
  • Performance varies by task. The same report said agents were much stronger at monitoring and communication work than at building evaluation, reinforcement-learning and agent environments, where success was far lower.
  • Human ideas still dominate. OpenAI said people originated nearly all of the key ideas behind GPT-6 Astra and GPT-6.2 Astra, with models helping more on implementation than on research direction.
  • Usage and spend are climbing fast. The company said median researcher agent spend reached about $2,400 a day at list prices in October, while concurrent use of multiple agents also increased sharply.

The bottleneck is no longer access

For a business owner, this is a workflow design story more than a model story. The practical mistake is assuming that if an agent performs well on one class of work, it can own the full process end to end.

In reality, the highest-risk failures often happen in the setup layer. An agent may summarize a customer issue correctly, yet make bad assumptions when building the next step, structuring data, or applying the same pattern across dozens of cases.

That means the advantage shifts to teams that separate low-judgment work from high-judgment work. Let AI handle the scanning, drafting and first-pass analysis. Keep humans in charge of approvals, exceptions and the parts of the process where bad structure creates expensive cleanup.

The OpenAI report also highlights a cost lesson. Usage can grow much faster than trust. If your team starts many agent runs in parallel without clear boundaries, spend rises before reliability does.

How Metomorph helps teams use agents where they hold up

This is where Metomorph is useful as an operating layer, not just another chat box. Imagine your team handling inbound customer requests that arrive with documents, urgency and edge cases.

You could capture every request through custom forms so the intake is structured from the start, instead of asking an agent to infer missing details from a messy email. That alone removes a class of avoidable mistakes.

From there, organization skills can package the repeatable part of the work: summarize the request, pull the relevant background, draft the internal handoff, and prepare the response for review. The process becomes consistent because the team runs the same defined skill, not a different prompt every time.

Then business rules can hold the risky step for a person when needed. If the draft includes pricing, legal language or a sensitive exception, the workflow can require approval before anything leaves the building. That's a practical response to the exact pattern OpenAI described: agents are useful, but not equally trustworthy across every task.

The bigger lesson is to design around uneven performance. Metomorph lets you turn the safe middle of a process into structured, reusable work while keeping judgment and approvals attached to the points that deserve human attention.

What to do this week

  • Break one recurring workflow into three parts: intake, repeatable processing and approval. You'll quickly see where an agent can help and where it still needs review.
  • Measure success by task type, not by one overall AI score. Drafting, monitoring and summarizing may be ready first while setup and exception handling lag behind.
  • Put the process into a form or reusable skill before scaling usage. A defined workflow is easier to improve than a pile of ad hoc prompts.
  • Pick one live process this week and add a human approval step at the highest-risk moment. That's usually the fastest way to raise trust without slowing everything down.

Explore Custom forms, Organization skills, Business rules in Metomorph.


Source: PPC Land, OpenAI's research agents now finish 46% of two-to-four-hour tasks alone (November 22, 2026).

Share this insight
in X

Keep reading

Turn your company context into answers.

Bring your tools, knowledge, team, and AI into one shared workplace.