Agents· 3 min
Agentic workflows
Instead of one big prompt, break the task into steps: reflect, plan, use tools, review — often via multiple LLM calls.
Key idea
Andrew Ng's four patterns: reflection, tool use, planning and multi-agent collaboration — together they beat single-shot prompting on hard tasks.
Example
A coding agent writes code, runs it, reads the error, patches it, and only returns once tests pass.
What to remember
- Small model + smart workflow can beat a large model + naïve prompt
- Latency and cost grow — design for parallelism
- Evaluation must be at task level, not per-call