This is a two-pass Claude Code assessment regarding the value of the default lenses under the v1 search API. A custom MCP tool exposing the lens_id parameter and a list of default lenses was used for the assessment. The second pass was an adversarial review to challenge the findings and methodologies of the first agent.
The original discussion started on the kagi-api Discord channel, and Margret asked me to repost it here with added detail. Considering the volume of feedback I've done my best to break it down for online readability.
First agent
Headline: only 3 of the 8 enumerated lenses earn their slot
| Verdict | Lenses |
|---|
| Valuable (hard to replicate with query operators) | academic, pdfs, kagi documentation |
Replicable with site: union (lens adds no value) | forums |
| Harmful (surfaces tag pages instead of question pages on specific queries) | programming |
| Low utility / misnamed | news 360 (no-op on fresh news), fediverse forums (returns Reddit + HN, not fediverse), usenet/archive (returns modern blogs, not archives) |
The programming finding is the most striking: it actively degraded a query that asked a specific technical question, surfacing StackOverflow tag-listing pages instead of question pages. Recommending it as an LLM-facing slug is a foot-gun.
The two misnamed lenses (fediverse forums, usenet/archive) don't deliver what their names imply — naming them as "always-available" invites the LLM to reach for the wrong tool.
Agent's recommendation
Trim the enumerated catalog in the Field description to three slugs: academic, pdfs, kagi documentation. Keep shareable IDs and full URLs as accepted forms (that's the durable value — user-curated stored profiles can't be replicated with operators). Keep the generic-behavior caveats (case sensitivity, silent fallback, workflow caveat, intersection/override precedence) since those apply to any slug value.
The agent estimates this removes 80 words from the Field description, eliminates two foot-guns, and concentrates the documentation budget on the lenses that earn it.
Second agent (adversarial of initial findings)
The adversarial round substantially shifted the picture. Three of the eight verdicts overturned.
Revised verdict table
| Lens | First agent | Adversarial review | Reason for overturn |
|---|
forums | drop | keep | Lens reaches Discourse instances, Facebook Groups, vendor forums (Swift Forums, discourse.nixos.org) that an LLM-written 3-domain site: union doesn't reach |
programming | drop | confirm drop | Narrow-utility band rather than "harmful across the board" — works for general/beginner queries, degrades on expert-niche queries; hard for the LLM to predict which |
news 360 | drop | keep | First agent's GPT-5 query was already news-saturated at baseline, masking the lens. On cold-news queries (Tigray humanitarian, Argentina labor reform) the lens shifts to actual news outlets (Al Jazeera, Reuters, argentinareports.com) |
fediverse forums | drop | confirm drop | Activates only on fediverse-native vocabulary queries (!community@instance syntax), not paraphrased queries — brittle in a way agents can't predict |
usenet/archive | drop | confirm drop | Doesn't reach groups.google.com or archive.org Usenet collections even on Usenet-native topic queries |
academic | keep | confirm keep | The 3-domain operator union returns preprints; the lens reaches peer-reviewed venues an LLM wouldn't enumerate |
pdfs | keep | drop | filetype:pdf actually outperforms the lens — the lens is a soft prior with site biases, not a strict PDF filter (returned GitHub markdown when asked for PDFs) |
kagi documentation | keep | confirm keep | Stable across multiple Kagi-self-referential queries (search, Orion, Assistant) |
Net change: first agent had 3 keeps / 5 drops; adversarial review has 4 keeps / 4 drops, with forums and news 360 moving keep, pdfs moving drop.
The methodological signal is the bigger finding
The reviewer's note that the value-assessment loop has not converged is the more important takeaway. Two rigorous rounds moved 3 of 8 verdicts. That tells us something about the parameter itself, not just our process:
- Lens value is query-shape-dependent. Whether a lens helps depends on whether the baseline is already saturated with the lens's target class. The LLM caller would have to reason about query shape to decide whether to engage the lens — which adds to the cost of using it.
- The enumerated catalog needs more than one query per lens to assess fairly. Single-query verdicts demonstrably mislead. The reviewer recommends 2 queries per lens with deliberate phenotype variance (general vs niche, saturated vs cold baseline).
- Doc-trimming pressure biases toward "drop it." The reviewer specifically flagged that running the value assessment in the same pass as drafting docs is methodologically unsound — drafting pressure creates a bias toward exclusion. The first agent's
forums and news 360 verdicts both swung that way.
