@hackproduct9: Stop making every node an LLM call. That single habit is why your agent breaks at scale. The question isn't "which framework." It's: which parts of this actually need a model, and which parts are just code you're paying a model to guess at? Here's a real graph. A customer asks: where is my order? 🔵 CLASSIFY — LLM. Which intent is this? ⚙️ FETCH ORDER — code + MCP. One API call. No model. ⚙️ LOAD POLICY — code. Read from cache. No model. 🔵 DRAFT REPLY — LLM + MCP. Writes the answer. ⚙️ CHECK — code. Is the tracking number actually in it? 🔀 PASS? — a router. Reads a boolean. Not a model. ⚙️ SEND — code + MCP. 👤 ESCALATE — a human. The escape hatch. 8 nodes. Only 2 of them are the model. 👀 The insight nobody says out loud: your router should be code, not a model. The moment you ask an LLM "what should I do next," you've made your control flow nondeterministic — and you can no longer reason about, test or replay your own system. CHECK produces a boolean. PASS? reads it. That's it. DETERMINISTIC WHERE YOU CAN BE. A MODEL ONLY WHERE YOU MUST BE. Two more rules on screen, and they matter as much: 🔁 The retry is bounded. No, retry, max 2. Unbounded retry loops are the number one way agents burn $4,000 overnight. 🚪 There's an escape hatch. ESCALATE stays greyed out for the entire animation. It never fires. That's the point of having it — a graph with no terminal human path doesn't fail, it just spins. The failure mode: teams draw the beautiful six-node diagram where every box is an "agent," ship it, then discover their p99 is 40 seconds and their bill scales with traffic in a way nobody modelled. Send this to whoever is about to make every node an agent. 📸 Screenshot the last frame — the typed graph plus the legend: NODE, EDGE, STATE, ROUTER, RECOVERY. Save it before your next architecture review. Follow @hackproduct — we turn scary AI-engineering concepts into things you can actually ship. ⚡ . . #AIengineering #softwareengineering #machinelearning #AIagents #systemdesign

HackProduct

Open In TikTok:

Region: US

Sunday 23 August 2026 19:01:31 GMT