Your Best AI Agent Isn’t the Smartest Model
LinkedIn post (anonymous poster; built on Andrew Ng’s agentic workflow) — view post
Core claim: an agent has amnesia without both loops and graphs — loops give it thinking, graphs give it memory, you need the two together. The advantage is the architecture, not the model.
The 4 workflows, step by step
- Reflection — the agent writes, a second prompt critiques, the agent rewrites. One self-review loop beats a smarter model that skips it.
- Tool use — search, code execution, APIs. A model reasoning with no tools is just guessing at facts it can’t check.
- Planning — break the task into JSON steps before running; when a step fails, the agent replans around it.
- Multi-agent — stop running one agent, run a team: one codes, one reviews, one tests.
Wiring it today
- Add one critique call after every generation — a 10–30% quality lift for about a day of work.
- Connect all four into a graph so agents share memory instead of re-reading transcripts. Hardest part is deciding what gets written, read, and pruned in that shared memory — a design problem before a code problem.
Industry takeaway
“The winners aren’t always the biggest, they’re the ones who build the right system around what they have.” Post ends with a funding pitch for GenAI Works (taking NVIDIA/Oracle/Google + 300+ AI brands to market; raising from $1,000 shares, deadline Aug 31) — treat that part as marketing.
Comments
Aleksei Mitrofanov (CPO): biggest gains rarely come from switching models — they come from designing a better operating system around the model; tool use + reflection loops give fastest value on internal products.