When Web-Connected Agents Keep Making Things Up, Don't Switch Models Yet: They May Be Living on Search Snippets

Santiago posted a video with a very specific topic: how to help agents like Codex, Claude Code, Cursor, and OpenCode get better results online.

In his reply, he included a link to the Apify MCP server and noted that this was a collaboration with the Apify team. The MCP server is positioned to let agents directly call 104K+ tools for web scraping, data extraction, and automation. Its MCP address is https://mcp.apify.com/, it supports OAuth or API tokens, and clients such as Claude Code, Codex, and Cursor have corresponding configurations.

The comments point to something more specific. One user, vexor4i, put it directly: give the agent clean context, remove noise from the query, use precise terms, verify sources yourself, don't trust the first result, and iterate quickly. Jon Cheney added a sharper point: half of the bad agent output he sees is poor search, not poor reasoning. Fix what the agent reads, and it will get noticeably smarter. Ashley Maloney was even more specific: search snippets are exactly where agents pick up confident nonsense; reading the full text can fix most of it.

Put these together, and the conclusion isn't complicated. Model capability certainly has limits, but many online tasks fail first at the input layer. To save tokens, search snippets compress web pages into a few sentences, losing qualifiers, dates, counterexamples, and original context. When an agent gets this secondhand information and then fills in the gaps, the output can be wrong in a very polished way.

So check the pipeline first: Are the query terms precise? Are you feeding it only snippets? Have you had the agent read the full page? Do you require it to cite sources? Do you allow it to retry quickly? TinyFish takes another path, putting TinySearch, TinyFetch, TinyBrowser, and TinyAgent under one API key. TinySearch and TinyFetch are free, and MCP connections are also provided.

Tools can save effort, but don't mistake tools for the answer. The real value of Santiago's video and the comments is that they break an agent's online performance into checkable stages: search sources, page reading, context cleaning, source verification, and iteration speed. Whichever stage cuts corners will eventually show up in the form of “the model is making things up.”

Image and video sources: Santiago's X post; Apify MCP server; TinyFish.