{
  "$type": "site.standard.document",
  "content": "---\ntitle: \"Agentic AI: LLMs with stones\"\ndescription: \"Agentic AI is about giving LLMs tools to act in the world without asking\n  first. The real shift isn't new capabilities---it's the removal of the human\n  bottleneck.\"\ntags:\n  - ai\n---\n\n> Sticks and stones may break my bones, but words will never hurt me.\n\nThere's a truth to that proverb, even if you feel (as I do) the temptation to\n\"well akshually...\" make several very valid points about how words _can_ be\nhurtful. For most of the Large Language Model (LLM) era, since the public\nrelease of ChatGPT in November 2022, we've been by turns amazed, disgusted and\nnow kindof \"meh\" about the way that LLMs can take the **words** we give them and\nproduce **more words** in response.\n\nWorking as I do as an\n[academic computer scientist](https://cybernetics.anu.edu.au/people/ben-swift/)\n(with a research background in AI) who regularly runs executive education\ncourses on AI for a diverse range of educated and intelligent folks, I'm getting\nmore and more questions about \"agentic AI\". And while definitions and\ndescriptions change pretty quick in this field at the moment, I want to\ndemistify some things about this term in particular.\n\nAgentic AI (as concieved and talked about in this present moment) is about\nconnecting LLMs (those \"pure\" input/output text sausage machines) to the world\nwith **tools** they can use to do stuff beyond just returning words in a text\nbox on a web page. To return to the \"sticks and stones\" aphorism above: agentic\nAI means giving an LLM a stone.\n\nHere's how it works in practice:\n\n- you put in a prompt as normal which is sent to the LLM\n- in addition to that prompt, though, the LLM is sent a list of tools that you\n  have access to (including human-readable descriptions of what they can do),\n  for example:\n  - `calendar`: add, update, or delete events in your calendar\n  - `weather`: check the weather forecast for a given location\n  - `fire-ze-missiles`: launch a missile at a target location\n- instead of only being able to respond with text, the LLM can\n  respond[^tool-call-llm] with a \"tool call\" instruction, to continue the\n  example:\n  - use tool `calendar` to \"add a meeting with John at 10am tomorrow\"\n  - use tool `weather` to \"check the weather forecast for Sydney next Tuesday\"\n  - use tool `fire-ze-missiles` to \"launch a missile at coordinates 40.7128,\n    -74.0060\" . it can say \"add a meeting to the calendar at 10am tomorrow using\n    the `calendar` tool\"\n\n[^tool-call-llm]:\n    The LLM needs to be specially trained to support this, but all of the main\n    ones do these days.\n\nIf the LLM requests a tool call, the user doesn't need to do anything; the\nsystem will use the tool as requested by the LLM and return the results (usually\nthe fact that this is happening is communicated to the user via some sort of\nvisual feedback in the interface, although this isn't a requirement).\n\nVocab-wise, this all started with OpenAI introducing \"function calling\" to GPT\nmodels in\n[June 2023](https://openai.com/index/function-calling-and-other-api-updates/).\nHowever it wasn't until late 2024 that the term \"agentic\" really took off,\ncoinciding with Anthropic's release of the\n[Model Context Protocol](https://www.anthropic.com/news/model-context-protocol),\na standardised and interoperable way for other parties (not just the LLM\nproviders) to create tools which all LLMs could use. That term is just riffing\non the \"agency\" sense of the world, where LLMs are given the means of acting in\nthe world. But agentic AI, tools, and tool/function-calling LLMs are all the\nsame general idea.\n\nIn the last couple of years there have been a proliferation of such tools. Some\nof them are really general, e.g. \"search the web for ...\". Some of them might be\nreally specific to your company, e.g. a tool that maps names to phone numbers in\nyour company's database. In this case they're useful because they're _not_\nLLM-powered (and so they don't just make stuff up if they don't know the\nanswer).\n\nSoftware developers (including me) in particular have found ways to use tools to\nhelp them write code. There have been many recent blog posts of various\ndevelopers describing how they set up their agentic AI (tool-calling LLM)\nsystems---from\n[Phil Schmid's \"Context Engineering\"](https://www.philschmid.de/context-engineering)\nto\n[Thomas Ptacek's \"My AI Skeptic Friends Are All Nuts\"](https://fly.io/blog/youre-all-nuts/)\nto countless Hacker News threads debating whether this is all just hype.\n\n## So is this a big deal?\n\nFrom a cybernetic perspective, this isn't quite as big a change as you might\nthink. Because even the original ChatGPT could \"do stuff in the world\" by\ntelling _you_ (the human user) to do it. Sometimes that was as benign as having\nyou copying text into an email. We'd still colloquially refer to this as\n\"answering my emails with ChatGPT\", but actually all ChatGPT was doing was\ngiving you **words** to type into your email client and hit \"send\". Sometimes\nthe LLM's words told us to do more life-impacting things, like\n[break up with your partner](https://www.vice.com/en/article/we-asked-chatgpt-how-to-break-up-with-someone/),\nor worse. Whenever LLMs are used by humans they have the (indirect) ability to\naffect the world.\n\nIn my opinion the best way to think about this shift isn't that LLMs can now\ninfluence the world; it's that now they can do it without asking, and this\ntightens the feedback loop. This means:\n\n1. first, there's now no longer a human in the loop (so now there's no human to\n   say \"hey, that's a dumb idea\" and refuse to do it)\n2. as a consequence of #1, LLMs can now run/iterate without intervention for\n   much longer (minutes, maybe even hours...)\n\nThe second point is the bigger deal (and it's a point that Anthropic, the makers\nof the Claude LLM which is one of the big players these days, makes in their\n[recent whitepaper about agentic AI](https://www.anthropic.com/engineering/building-effective-agents)).\n\nHumans were always a) able to do things in the environment, and b) the\nbottleneck in any LLM system (time-wise, at least). But by gaining the former\ncapability, agentic AI removes the latter bottleneck.\n\nSo if you're going to allow your LLMs to use tools, you **must** be certain that\nyou're comfortable with what things the LLM can do with them. Both in theory,\ni.e. in the sense of what's possible, but also in practice, i.e. through testing\nthe way that your particular LLM tends to use your tools given specific prompts\nor other context. How exactly you do that is a topic for another blog post (just\nkidding, it's a much bigger question than that that depends on a whole bunch of\nthings). But I think that's the right question to be asking when it comes to\nagentic AI.\n",
  "createdAt": "2026-05-13T23:14:44.058Z",
  "description": "Agentic AI is about giving LLMs tools to act in the world without asking first. The real shift isn't new capabilities---it's the removal of the human bottleneck.",
  "path": "/blog/2025/07/17/agentic-ai-llms-with-stones",
  "publishedAt": "2025-07-17T00:00:00.000Z",
  "site": "at://did:plc:tevykrhi4kibtsipzci76d76/site.standard.publication/self",
  "tags": [
    "ai"
  ],
  "textContent": "Agentic AI is about giving LLMs tools to act in the world without asking first. The real shift isn't new capabilities---it's the removal of the human bottleneck.",
  "title": "Agentic AI: LLMs with stones"
}