AI agents do more of the work in model development, but humans still make the decisions

What changed

A research team analyzed 769 task logs from building their own AI model to see how much work AI agents actually do versus humans. AI agents proposed up to 55 percent of the methods used in development tasks. Despite this, humans made over 85 percent of the final decisions. Notably, roughly one-third of the tasks would not have been attempted without AI agent involvement.

Why builders should care

The finding challenges assumptions that AI agents are driving full automation in model development. While they generate a significant share of ideas and methods, humans still hold the reins on which approaches get implemented. This means human judgment remains critical for steering AI workflows and managing risks. Relying solely on agent activity metrics can be misleading when evaluating autonomy or efficiency gains.

The practical takeaway

Operators and AI builders should recognize that AI agents act more like collaborators or amplifiers than independent decision-makers. Agents expand the range of considered options but don’t replace human expertise in selecting and refining solutions. Planning for AI-assisted development needs to include clear decision-making roles for humans. It also opens opportunities to scale exploration without losing control but requires monitoring to avoid blind reliance on agent suggestions.

What to watch next

The balance between AI agent contributions and human decisions will shape how workflows evolve. Watch for tools that improve transparency and traceability of agent proposals versus final choices. Expect ongoing discussion about which tasks can be safely handed over to agents and which demand human discretion. The next step is refining collaboration frameworks to harness agent speed while preserving human judgment.

AI Quick Briefs Editorial Desk