OpenAI’s web search guide (read 6 October 2026) draws this line. Non-reasoning web search hands the query to the search tool and relays the top results, with “no internal planning”. Agentic search lets a reasoning model search inside its chain of thought, analyze results, and decide whether to keep searching, so it takes longer. Deep research runs the loop for several minutes, often across hundreds of sources. Google’s AI Mode announcement (20 May 2025) calls its version query fan-out: the question is split into subtopics and many queries run at once.
How does agentic search differ from regular AI search?
Three things change. The agent, not the user, writes the queries, often several in parallel, and query rewriting turns a vague question into searchable phrases. The agent judges results before it answers and can open a full page instead of trusting a snippet. And it decides when to stop, so cost and latency vary per question. Anthropic’s web search tool docs (read 6 October 2026) put a simple factual query at one to three searches and comparative research at ten or more, priced at $10 per 1,000 searches plus tokens, with a max_uses cap per request. Over a private index the same loop is agentic RAG, often pairing semantic search with keyword matching.
What have we learned running agentic search at Soba Labs?
Our content research runs on agentic search: several research agents search the live web in parallel, and a coordinating agent works only from what they return. We treat a search result as a lead, never as the fact. Three rules came out of that work:
- Cap searches per agent. Our scheduled runs share 200 web searches per session; one test run spent all 200 on research and could not run its final check of live results. Each of four parallel research agents now gets 20 and returns what it has at the cap. Opening a page is not a search.
- Search finds, fetch reads. A changing value, such as a price or limit, is read on the live page beside its date or version label, never from a snippet. One email search tool we use drops a thread’s newest messages, and a check of our sent mail missed a reply sent four minutes after the inbound, so we now open the full thread by ID.
- Keep search and the record apart in tool design. In krs-mcp, our open-source server for the Polish company register, one tool resolves a name to a registry number and others fetch the official extract, each result carrying its source URL and retrieval time.
The wider rule is in what AI summaries drop.
When should a team use agentic search?
It earns the extra calls on questions that span sources, need corroboration, or carry a high cost for a confident gap. A narrow question with a strong index needs one retrieval pass, not an agent loop. For long investigations, deep research agents add explicit source review and synthesis, and a planner-executor split keeps the search plan inspectable.