Blog

The built-in web search inside coding agents is blocked from Reddit at the crawler level — not rate-limited, blocked. That's the one source where people say what actually happened. A structured endpoint reopens the door.

Your users file tickets. Everyone else complains on Reddit and never tells you. Here's an agent that reads the second group and reports what's actually breaking.

We ran the same research question through Claude Code's built-in WebSearch and WebFetch, then through Serply. They tied on the academic paper. One hard access boundary decided everything else.

Reddit's per-call rate is cheap. Getting to it — approval, tiers, and rate limits — is where most agent projects stall. An honest look at both paths.

Upvotes measure agreement, not accuracy. Four checks — recency, account history, dissent, and cross-source — that turn a thread into evidence instead of a vibe.

A big thread is a deep tree, not a list. Flattening it naively wastes most of your budget on replies nobody upvoted. Here's the pruning that keeps the signal.

Every other source tells your agent what a product claims. Reddit tells it what happened to the people who bought one. That gap is why agents keep reaching for it.

A browser agent pays to re-read the same page on every step. Where the money actually goes, why open source doesn't fix it, and which of your tasks never needed a browser.

Indexing your documentation into a vector store means re-indexing it forever. Site-scoped search reads the current version every time.

Footnote-style links at the end of a paragraph don't get checked. Making provenance visible at the sentence level changes what users trust.