At CoRL, Berkeley's Tom Silver Dissects "Agents Are Too Slow" Into Three Distinct Critiques

During CoRL 2026, Berkeley robotics researcher Tom Silver posted from the conference floor to untangle a criticism heard throughout the event: many talks and discussions argued that "agents are effective for robot decision-making, but too slow and not reactive enough" to justify approaches like VLA/WAM. Silver contends that this sentence conflates three fundamentally different issues, and that they should be separated into three distinct critiques and discussed on their own terms.

Confirmed

• The poster is Tom Silver, a robotics researcher at Berkeley; the post was made on November 11 from CoRL 2026.
• The core claim: this popular criticism actually bundles three different concrete problems that need to be treated separately (the post's detailed breakdown of each critique was truncated in this cluster of materials, so the specifics are unavailable).
• The post was shared multiple times that day, all pointing to the same original text with no additional new information.

Why it matters

"Agents are too slow" is one of the key arguments currently used in robot learning to justify end-to-end VLA/WAM approaches. If the criticism itself is conceptually muddled, then methodological choices derived from it (such as abandoning explicit agent planning in favor of reactive policies) need to be reexamined. Silver's breakdown gives the community a sharper framework for reasoning about the relationships among latency, reactivity, and decision-making effectiveness.

2026-11-11 ~ 2026-11-11 · 5 related posts

• CoRL researcher unpacks "agents are too slow for robots" into three distinct critiques — tomssilver · 2026-11-11

4 near-duplicate retellings: tomssilver · tomssilver · tomssilver · tomssilver