Jargon:Agentic
Say it: AY-jen-tik · 1 min read
Say this at the party
“It means the AI doesn't just answer you. It makes a plan, uses tools and keeps going until the job is done.”
In plain English
A normal chatbot is a very clever reply machine. You ask, it answers, it stops.
An agentic AI is given a goal instead of a question. It works out the steps, does them one by one (searching, writing, clicking, running code), checks how it went, and fixes its own mistakes along the way.
Think of the difference between asking a friend for directions and asking a friend to drive you there.
Where it came from
The idea of AI "agents" is decades old. The word "agentic" took off in 2024, when Andrew Ng started talking about "agentic workflows" and showed that letting a model plan, check and redo its work could beat simply using a bigger model.
See it in action
Ask Claude Code to "fix the broken sign-up form". It reads the code, finds the bug, changes it, runs the tests, sees one fail, fixes that too, and tells you when it's done.
What people get wrong
- Agentic doesn't mean the AI has wants or goals of its own. It means it can take several steps without you approving each one.
- Lots of products call themselves agentic when they are a chatbot with one button. If it can't take more than one step on its own, it isn't.
Go deeper
- PaperYao et al., 2022ReAct: Synergizing Reasoning and Acting in Language ModelsThe paper behind most agents today: think a step, act a step, look at the result, repeat.
- PaperWang et al., 2023A Survey on Large Language Model based Autonomous AgentsThe big overview, if you want the whole field in one place.
- ArticleAndrew Ng, The Batch, 2024Four AI Agent Strategies That Improve GPT-4 and GPT-3.5 PerformanceWhere "agentic workflows" went mainstream.
- GuideAnthropic, 2024Building effective AI agentsThe clearest practical guide. Its advice: start simple.
Where it comes up
- Your new job is managing the AIsAnthropic studied its own staff. Agents now do six hours of work for every human hour, and people spend more of their time on “meta-work”.Our take →
- A production line for content, not one big promptA YouTuber split his content work into seven agent stages and cut two or three days down to about five hours.Our take →
- One developer, two dozen agents, zero alignmentAI agents are built for one person working alone. Real work is done in teams, and that's the new bottleneck.Our take →
- Stop prompting, start managingEthan Mollick's guide to picking an AI now that they do whole tasks: think about the model, the app and the harness.Our take →
- Design the loop, not the promptAn agent is just tools in a loop toward a goal. Simon Willison shows how to set one up so it can safely work things out.Our take →
Related jargon
Drafted with AI, checked by the AI Council (Claude, ChatGPT and Gemini) and signed off by Phin.