Andrew Ng on Four Agentic AI Design Patterns

SUMMARY

In 30 seconds

Matthew Berman introduces an excerpt of Andrew Ng’s 2024 talk on agentic workflows. Ng discusses four design patterns and the tradeoff between model quality, speed and token costs.

1A talk on agentsBerman introduces Andrew Ng and the excerpt.
2Four patternsNg names reflection, tool use, planning and multi-agent collaboration.
3Speed and costNg considers whether faster tokens from a smaller model can improve some applications.

Takeaways

Ng identifies reflection and tool use as practical agent patterns.
He describes planning and multi-agent collaboration as emerging approaches.
Faster, lower-quality models may be useful in some workflows, depending on costs and task performance.

TopicsAI agents Agent engineering Multi-agent systems

FULL TEXT

Read the edited text by chapter

Show transcript

Transcript only

Agent design patterns

Ng introduces four patterns after Berman’s introduction.

Matthew Berman introduces a talk by Andrew Ng about practical approaches to AI agents. Ng says he wants to make the discussion concrete by describing recurring design patterns.

SLIDE 1 / 1Four agentic reasoning patternsThe slide lists reflection, tool use, planning and multi-agent collaboration.

Ng identifies reflection and tool use as useful patterns, while describing planning and multi-agent collaboration as more emerging. He says these methods can sometimes produce surprisingly strong results.

Model speed and cost

He discusses a possible tradeoff between token speed, model quality and cost.

Ng suggests that generating tokens quickly from a somewhat less capable model might work well for some applications compared with a slower, stronger model. He says the cost of those tokens also matters; this is a conditional observation, not a universal benchmark result.