Agent Context Is Longer Than Ours Now—Debugging Can Only Be Handed to Agents

Sarah Wooders asked on X: when an agent's context length exceeds our own, how do we debug them?

Sydney Runkle reposted it with one line: the reality now is that we need agents to debug agents.

The question isn't new; the timing of putting it on the table is interesting. Over the past year, agents have gone from single-turn calls to tasks that run for tens of minutes or even hours, with traces routinely tens of thousands of tokens. Humans can't read them all. And if we can't read them all, the act of 'review' exists in name only.

The replies are basically a grassroots list of proposals.

Alex Strick van Linschoten said he's been thinking about observability for agent swarms and shared a tool demo. His judgment is very grounded: without visualization tools, or without an agent to explain things on a person's behalf, it's hard for people to keep up with what's happening in a swarm.

Sreeuday wants a button called shrink this failure. Reduce the failure to a minimal context, shrink it until a human can understand it. This framing is smarter than 'add more logs': the direction is to subtract information, not add it.

Steven Zimmerman said decomposition. Abinash Karki said build a system and periodically or randomly spot-check; he called this the limit of his game theory.

suka's reply was the most pointed: it's like reviewing a 4,000-line PR, scanning the beginning, and clicking approve.

Dave Birnbaum offered a more engineering-minded approach: manually do 5% to 10% of the work, then use an outer loop to compare human output with agent output and generate suggestions for modifying the harness.

Michael Hogers said use more agents. Theodore Gavriilidis said an agent's context being longer than a human's sounds fine.

Aman asked just one thing: do we need to?

That line is worth pausing on. Debugging presupposes that someone can judge right from wrong. If an agent's reasoning chain becomes too long for a human to verify, the meaning of the word 'debug' changes. It is no longer about finding the cause of an error; it is about finding which step's output deviated from expectations. The former requires understanding; the latter only requires comparison.

Sydney's line 'need agents to debug agents' sounds like surrender, but it is actually describing a delegation chain that is taking shape: humans set the standard, agents produce, agents check, humans spot-check. Each layer retreats one step further down. At the end, the only thing left for humans to do is define what counts as failure.

That job is not easy. Defining failure is harder than fixing failure, and fewer people are willing to do it.