What Is Agentic RAG? An AI That Picks Its Own Tools

My AI agent answered the question every search-based AI got wrong. Then it froze. This is round four of the RAG Olympics. Same 567 business documents, same 30 questions and same AI brain as the last three episodes. Only the memory changes. Today: the Detective, which uses agentic RAG. Instead of one fixed way to look things up, it can think, pick a tool, read what comes back and decide again. Four tools: search the filing cabinet, open the spreadsheet, read a whole document, or ask the map (GraphRAG). Nobody tells it which to use. The test business is Grand Union Lettings, a made-up letting agency. It is demo data, not a client. Where it beat every search contestant: "If we served a rent increase notice at Flat 9, 125 Kensal Basin Close on 5 Oct 2026, what's the earliest the new rent could start?" It checked the spreadsheet (rent due on the 1st, last increase 1 June 2025), read the whole Office Handbook for the rent review rule and answered 1 January 2027, in 11.8 seconds for 12 cents. Every search contestant: 0 out of 6. The judgement call: "Which landlord relationship is most at risk right now, and why?" Nine steps, 31 seconds, 99 cents. It flagged one landlord with three complaints since April and two jobs still open. Told he'd been talking to another agent, it found four properties at risk. It didn't trust the spreadsheet. The spreadsheet said 7 gas certificates expire in October, but one was renewed last month. The Detective checked and said 6. With the spreadsheet disconnected, it read the register as a document and still said 6. Then it froze. On a simple rent lookup it got three steps in and hit "Max iterations (10) reached", sending the same prompt 8 times in a row. 5 of 30 answers never came back. That cost 10 of 18 points and $3.53 for no answer. The cause is not proven yet (it may be the n8n wrapper or the thinking setting). The verdict: Speed-Reader: 57 Detective: 43 Plain search: 35 Connector: 30 Guesser: 6 Librarian: 5 Second place, and the most expensive way to get there: 30 questions cost $13.93, against my estimate of $4 to $5 42 cents per answer, about 70 times the Librarian's Typical answer 17 seconds, slowest 73 seconds Fast enough for a phone call: 0% It re-reads the whole conversation on every step One tool vs four: on the same 17 questions (out of 34), the Librarian scored 1, the Detective with search only scored 14, and the Detective with all four tools scored 21. Just letting it keep searching took the score from 1 to 14. Credit ran out after 17 questions, so treat that as a direction, not proof. What it means for your business: 1. Agents win on multi-step questions: rules, judgement, "check the newer document". 2. Budget per answer, not per month. Ours cost about 3 times my estimate. 3. The plumbing can fail silently. Test the wrapper, not just the model. 4. Let it search more before buying anything fancy. 5. Read-only access. An agent that can change records is a riskier build. Honest limits: one official run (an agent's choices change from run to run), the cause of the freeze is not proven, the control ran 17 of 30 questions, the map ran on a laptop through a tunnel, fictional documents, 30 questions, and not legal advice. Next: the final. 5 AI memories, 30 questions, 1 winner. Accurate, fast, cheap and ready for a phone call: pick three. JARVIS needs a brain that can answer the phone. Comment your question. The best one goes to the winner, live. This is Project JARVIS, Act 2, Episode 4, where I build an AI employee in public. Last episode:    • I Built a Map of My Whole Business. The AI...   Chapters 00:00 Agentic RAG: an AI that picks its own tools 01:00 The scoreboard so far 04:39 Meet the Detective 06:36 The four tools it can use 08:41 Watch it choose: 183 tool calls 09:37 The pets question: 35 calls 10:53 The question search couldn't answer 12:02 The judgement call 13:05 The pets question, again 14:19 Take away its spreadsheet 15:29 Then it froze 16:48 The verdict: second place 17:29 The bill: $13.93 18:20 One tool vs four 19:34 Detective vs the rest 21:19 What this means for your business 23:07 Honest limits 28:07 Next: the final I'm Uwa, founder of Autore. We build AI systems for small businesses, and white-label builds for agencies and consultants whose clients want something they can't build. Got a client asking for an AI build? Paste the request into https://autore.ai/scope-check and get the scope and tier free, or book a call: https://cal.com/uwa-autore.ai/ai-oppo... Keep shipping. #agenticrag #RAG #aiagents #aiforbusiness #ProjectJARVIS