Andrew Ng outlines four practical steps for building agents from loops to graphs
theomitsa · x · 2026-07-24
A repost of Andrew Ng’s PDF lays out four agentic workflow steps for moving from simple loops to graph-like systems.
- Reflection: let the agent write, have a second prompt critique it, then rewrite — a self-review loop can outperform a smarter model with no feedback.
- Tool use: add search, code execution, and APIs; otherwise the agent is mostly hallucinating.
- Planning: ask the agent to break the task into JSON steps before execution and replan when a step fails.
- Multi-agent: split roles across a team — one writes code, one reviews, one tests.
The post frames these as practical wiring patterns for building better agents today, starting with a critique step and then adding more structure as tasks get harder.
Related event: Evolution of AI Agent Architecture: From Loops to Graph Engineering(8 posts)→
More from coding & agent
- Team-level AI agents: where should shared context and history live? — Al_Grigor · 2026-09-11
- Chaining dependent MCP tool calls: no rollback, duplicate risk — agentrsdg · 2026-09-11
- DeepMind-led paper makes design docs the source of truth, code disposable — SMART regenerates in 1.5-3h for ~$100 — Roger_M_Taylor · 2026-09-11
- Agent-built classifier labels 192k docs for $0.70 vs $13-26 with frontier LLMs — vanstriendaniel · 2026-09-11
- MathModelAgent gains traction: auto-solves math modeling and writes a submission-ready paper — jihe520 · 2026-09-11
- alphaXiv open-sources OpenResearch to run parallel research agents with any model — alphaXiv · 2026-09-11