New term: agentic

2024-03-21 · Publisher : Andrew Ng · Domain : Artificial Intelligence

Describes a system where a language model does not answer in one pass but works through steps: it plans, uses tools, reviews and revises its own output, alone or with other agents.

### Definition An **agentic** workflow makes a language model work in several passes instead of a single direct (zero-shot) answer. The model plans, looks up information, writes a draft, critiques it and revises it, calling tools or other agents when needed. ### Origin Andrew Ng (DeepLearning.AI, AI Fund) laid out the idea on March 20, 2024 in his letter in *The Batch*, then summed it up on X on March 21, 2024: "I think AI agentic workflows will drive massive AI progress this year — perhaps even more than the next generation of foundation models." He expanded on it in his talk "What's next for AI agentic workflows" at Sequoia Capital's AI Ascent, published online on March 26, 2024. ### What the term covers - On the HumanEval coding benchmark, Ng reports 48.1% for zero-shot GPT-3.5, 67.0% for GPT-4, and up to 95.1% for GPT-3.5 wrapped in an agent loop. - Four patterns: **reflection** (the model critiques its own work), **tool use** (web search, code execution), **planning** (multi-step plans), **multi-agent collaboration**. - Ng likens zero-shot prompting to writing an essay in one go, with no backspace allowed. ### Criticism and limits The adjective quickly became a sales argument applied to very different products. Every loop adds cost and latency, and autonomous planning remains less predictable than reflection or tool use. ### Key takeaway "Agentic" describes a way of orchestrating a model, not a kind of model: value comes from the working loop as much as from the model itself.

Tags : Vocabulaire, Agents, Workflows, Conception