The linkedin_category filter on fetch-prospects accepts only the exact strings from LinkedIn's canonical industry taxonomy, but the current tool description ("filter prospects by LinkedIn industry category") doesn't enumerate them or call out common-but-invalid aliases. LLM agents reasoning over the schema reliably guess plausible labels (e.g. "Computer Software") and hit a 422 round-trip before recovering.
This is a documentation-only fix — no behavior change — but it eliminates a recurring class of agent failure on the first call.
Proposed fix
Before:
linkedin_category: filter prospects by LinkedIn industry category
After:
linkedin_category: filter prospects by LinkedIn industry category. MUST be
one of the exact LinkedIn industry strings — see
https://developers.explorium.ai/categories for the canonical list. Common
inline examples: 'Software Development', 'Technology, Information and
Internet', 'Information Services'.
The parenthetical negative example is optional but high-value — it directly preempts the failure mode we observed.
For convenience, here's the suggested_fix payload our evaluation pipeline auto-generated against this run (verbatim, as a sanity check on the prose above):
{
"kind": "mcp_description_snippet",
"target": "fetch-prospects.filters.linkedin_category.description",
"anchor": "turn 2: 422 Unprocessable Entity, 'Computer Software is not a valid LinkedIn category'",
"rationale": "The agent guessed a plausible-but-invalid category and hit a 422. Enumerating the common valid values (and explicitly noting common wrong aliases) in the tool description prevents this guess-and-fail loop and saves a wasted round-trip."
}
Surfaced by automated transcript analysis at bren.run — we evaluate agent runs against MCPs and propose patches for the failure modes we see. Happy to share more runs that hit similar patterns if useful.
The
linkedin_categoryfilter onfetch-prospectsaccepts only the exact strings from LinkedIn's canonical industry taxonomy, but the current tool description ("filter prospects by LinkedIn industry category") doesn't enumerate them or call out common-but-invalid aliases. LLM agents reasoning over the schema reliably guess plausible labels (e.g."Computer Software") and hit a 422 round-trip before recovering.This is a documentation-only fix — no behavior change — but it eliminates a recurring class of agent failure on the first call.
Proposed fix
Before:
After:
The parenthetical negative example is optional but high-value — it directly preempts the failure mode we observed.
For convenience, here's the
suggested_fixpayload our evaluation pipeline auto-generated against this run (verbatim, as a sanity check on the prose above):{ "kind": "mcp_description_snippet", "target": "fetch-prospects.filters.linkedin_category.description", "anchor": "turn 2: 422 Unprocessable Entity, 'Computer Software is not a valid LinkedIn category'", "rationale": "The agent guessed a plausible-but-invalid category and hit a 422. Enumerating the common valid values (and explicitly noting common wrong aliases) in the tool description prevents this guess-and-fail loop and saves a wasted round-trip." }Surfaced by automated transcript analysis at bren.run — we evaluate agent runs against MCPs and propose patches for the failure modes we see. Happy to share more runs that hit similar patterns if useful.