Watch on YouTubeWatch on Instagram
About this video
A small AI on my laptop made up a phone number. Put the same model in a loop with tools, and it found the real one.
What actually makes an AI an agent: think, act, observe, repeat. I measured how much the loop matters, where it gets stuck, and what thinking and a bigger model fix.
Episode 12 of Learn AI in Under 3 Minutes, level 3: AI agents.
Measured on a 2020 MacBook Air (M1, 16 GB) with llama.cpp and Qwen3.5 (open weights, Feb 2026) at 4 bits (Q4_K_M, unsloth GGUFs). The test: a made-up airline's baggage help desk, five small files (passengers, flights, bag scans, airports, policy), 20 customer questions in 4 tiers by how many lookups they need (1, 2, 3, and 4+ with a sum), graded by program against answers computed from the files. Tools: list_files, read_file, search, calculator. Setups: no tools; one round of tools, then answer; the loop until it answers in words (cap: 12 model calls); the loop with thinking on (temperature 0.6, otherwise greedy). Qwen3.5-2B: 0, 3, 9 (3 hit the cap), 14 (16 on a second seed). With no tools it made up 15 of 20 answers; with one round, 17 of 20 final replies were another tool call. Qwen3.5-4B: 0, 3, 14, 19. Qwen3.5-9B (4-bit, 5.7 GB): no tools 0, loop 14. One machine and 20 questions: an example, not a benchmark. Traces on screen are the models' real steps, shortened.
Sources: Yao et al., "ReAct: Synergizing Reasoning and Acting in Language Models" (2022, arXiv:2210.03629), the think-act-observe pattern. Anthropic, "Building effective agents" (2024). Qwen3.5 model cards (Alibaba Qwen team). Meta's acquisition of Manus (announced December 2025).
0:00 It made up a number
0:27 What an agent is
0:41 Does the loop matter?
1:04 How loops fail
1:24 What still breaks
1:50 Bigger brain, same loop
2:18 My take
Music: Aitech by Kevin MacLeod, CC BY 3.0, via Wikimedia Commons (https://commons.wikimedia.org/wiki/File:Aitech_(ISRC_USUAN1100336).mp3)
Follow @PranavInnovates for the tech story that matters, every day.
What the video says
I asked a small AI on my laptop for the phone number of the desk holding a lost bag. It answered instantly. That number is made up. Same model, same question. But now it can open the help desk's files, and it keeps going until it knows. It lists the files, reads the airports, then thinks: which airport is her bag at? So it reads the passenger list, the flights, the bag scans. Held in Dubai, missed connection. Six steps, and the real number. Nothing was retrained. That's all an agent is: a model in a loop. Think: pick the next step. Act: use a tool, like search or a calculator. Observe: read what came back. Then round again, until it can answer. To measure it, I built a made-up airline help desk: five files, twenty customer questions, from one lookup to four and a sum. No tools: zero right. It made up fifteen answers. One lookup, then answer: three. And seventeen times it asked to look again. The full loop: nine. The loop with thinking switched on: fourteen, and sixteen on a second run. Without thinking, it gets stuck. Three questions hit my twelve-step limit. Here, it finds the bag tag, sees the bag held in Dubai, then searches for bag desk, gets no lines found, and runs the exact same search eight times. Thinking is the step that notices what failed. And every agent needs a step limit. Quick quiz. Even with thinking, one kind of question beat it. One lookup? Three lookups in a row? Or working out a claim from the rules? The claims: two right out of eight. Three lookups: eight of ten. One lookup: all ten. Amara's bag was four days late, and the rule says pay at most six days. It paid six: five hundred twenty-nine euros, not three hundred sixty-three. Bigger models help too. Qwen three point five, four B, in the loop: fourteen. The nine B: also fourteen. The four B with thinking: nineteen of twenty, every claim right. The one it missed: Ana's bag. It assumed the bag was on her booked flight, and never opened the scans. It was rebooked onto another one. Coding agents like Anthropic's Claude Code, OpenAI's Codex and Manus run this same loop, with more tools and much longer runs. My take: an agent is a model, tools and a loop. Give it a step limit, log every step, and check the answer, not the confidence. This week: give a coding agent one small job, and read its trace: think, act, observe. Next: tool calling, how a model actually presses the buttons. Episode thirteen.