Theo stops queueing agent messages, steers mid-run instead
Two engineers say interrupting a running agent is no longer the mistake it used to be.
Theo says he does not like queueing messages to a running agent. His argument is that modern models are smart enough to take a steer message mid-task and not get distracted by it, so there is no reason to sit on a correction until the current turn finishes. He adds that the default is configurable in T3 Code.
Kent C. Dodds replied that the question does not come up for him, because the tool he uses as his main driver does not let you queue messages at all. He says giving it a pile of work and then interrupting it is not a problem, even when the new instruction has nothing to do with the old one. His guess at why is that the tool delegates everything to subagents.
It's not a problem to give it a bunch of stuff to do and interrupt it. It just figures it out even if the subjects are completely different.
Why the habit exists
Queueing a message means typing your next instruction while the agent is still working and letting the client hold it until the turn ends. It became a default in agent UIs for a practical reason, which is that interrupting a model mid-task used to cost you the thread. Both of these are single-engineer reports on their own setups rather than anything measured, so treat them as such.
What makes it worth a minute of your attention is that it is another habit built around a limitation nobody has rechecked lately. That is the same shape as dax arguing that custom agent setups fix problems that no longer exist. The workaround outlives the thing it worked around, because nobody wants to spend a run finding out whether the fix is still load-bearing.
For working engineers the cheap test is one keystroke. Next time you catch the agent heading the wrong way, send the correction immediately instead of queueing it, and see whether the run recovers or derails. If Theo and Dodds are right you save a full turn every time you spot a mistake early. If they are wrong on your stack you lose one run and you learn something about your harness.
Have the good opinion, not the loud one
One email a day. Under three minutes. What working engineers found out about coding with AI.