Skeleton keys weren’t “smart keys”. They didn’t have mysterious powers which allowed them to bypass locks, and they didn’t use magic or superstition to will an open lock into existence. Skeleton keys worked – in older, antique locks – because the locks shared underlying mechanisms which, at a fundamental level, could be operated by a sufficiently ground down key.
Large Language Model-based AI is a digital skeleton key. The systems are more complex, and the fundamentals not quite as obvious, but the metaphor of operating at a simplistic level to “open doors” is still useful.
Much of the work that happens on a computer is mediated by software. Documents are files; spreadsheets are structured data and formulas; websites are code, protocols and databases; email, calendars, project-management platforms and customer records are accessible through interfaces and APIs. Even applications without a friendly API can increasingly be operated through browsers, terminals or computer-use systems, including via Model Context Protocol (MCP) servers with AI.
Put a sufficiently capable language model inside an environment that lets it read files, write files, execute code and inspect the results, and it can potentially operate across any code-accessible system. Like the skeleton key, there’s no magic here. But once the model is authorised to use the machine, the boundaries between “writing assistant”, “data analyst”, “research tool”, “programmer” and “administrative assistant” begin to look artificial.
In common terms, this is the work of “AI agents”. The term is fraught with hype and buzzwordiness, like most other terms associated with AI, but it’s possible to look beyond that and land on simple definitions. Software engineer Simon Willison has probably the most concise:
An LLM agent runs tools in a loop to achieve a goal.
Simon Willison – I think “agent” may finally have a widely enough agreed upon definition to be useful jargon now. 18th September 2025
I’ve also attempted a taxonomy of AI agents, and Willison’s core premise – running tools in a loop – is at the heart of every form of “agent”.
It’s also a key feature of the “If You Know, You Know” series of posts I recently compiled into a free eBook, which focuses on the discoverability of AI features and the use of AI beyond chatbots.
You can grab that free eBook on the website at https://leonfurze.com/iykyk/
This article follows a similar trajectory, treating the “tools in a loop” aspect of AI as the digital skeleton key which, applied well, can unlock much more of the potential of LLMs than using them as simplistic writing tools.
From Deterministic to Probabilistic (and back again)
Traditional software is intended to be deterministic. Given the same code, inputs and environment, it should produce the same result. Floating point errors and cosmic rays notwithstanding, when you build software it’s intended to just do stuff.
Large Language Models, however, are probabilistic: they work by generating the most probable likely output, based on training data and other inputs.
The probabilistic nature of LLMs is what allows them to be so flexible in generating fluent-seeming text and other media. It’s what allows an application like ChatGPT to respond in sometimes surprisingly human ways. It is also the cause of negative effects, including AI hallucinations. While it’s possible to control the “temperature” (randomness) and “top-p” (how wide a net the model casts for its next word) of an LLM, you’re still essentially tinkering with a range of probabilities.
Tool use, however, changes the process. A tool-using AI agent does not calculate your figures by guessing; it writes a Python script that calculates them, and Python is deterministic. The AI agent doesn’t hallucinate the contents of files; it extracts the data and runs a search tool over them. It doesn’t guess what a webpage says; it fetches the page with a web search tool, and perhaps even uses image recognition to “view” the page like a human user might. The language model operates deterministic components like programming languages, command line tools, and APIs, rather than just generating plausible words.
Holding the Skeleton Key
A product like Claude Code in the terminal makes the process more visible. In Claude Code, you can watch as the model runs bash commands, writes code, and changes files directly on the device. Unfortunately, if you’re not a “computer person”, using the terminal can feel rather like falling into the Matrix, and comes with the uneasy sense that you’re about to break the computer (which might also be true).
Most users find the terminal uncomfortable, and the AI developers know it. Anthropic built Claude Cowork explicitly because, for many people, the terminal remained an intimidating black box, and they wanted to extend Claude Code’s agentic capabilities into the chat interface people already knew. OpenAI wraps the same class of capability in an app called Codex. Microsoft, in March 2026, launched Copilot Cowork, built in partnership with Anthropic on the back of a reported US$30 billion Azure compute deal, folding the agentic features into the Microsoft 365 products enterprises are more familiar with.
Wrapping the agentic features of Claude Code and similar applications in the more familiar chatbot interface of Cowork and Codex seems like a good idea, and also allows for additional layers of security and assurance. Even though Claude Cowork has the same permissions and file access as Claude Code, it feels qualitatively different to use.
Unfortunately, as I discussed in the IYKYK series, the chatbot interface also introduces some serious limitations. Anthropic’s June 2026 Economic Index begins by observing that ordinary chat transcripts no longer capture the whole picture because Claude Code and Cowork have produced more long-running agentic sessions. Yet, in its analysis of Claude chat and Cowork conversations, the most common outputs were still “explanations”, at 17 per cent; documents and reports, at 15 per cent; and guidance, at 11 per cent. Code and technical work accounted for about one sixth.
The July Cowork analysis makes the pattern even more obvious. Anthropic classified 1.2 million sampled Cowork sessions from May 11–31 into 20 categories. Business process and operations accounted for 33.4 per cent. Content creation and copywriting accounted for another 16.4 per cent. Software development, the thing agentic AI really built its name on, was just 8.7% in these Cowork platforms.
These uses remain adjacent to the established chatbot mental model. They involve producing, condensing and rearranging information, or treating LLMs as “answer engines”, like a super-powered Google search. Unfortunately, they don’t really surface the capacities of a system that can also operate software, build tools, connect processes and execute whole workflows.
In short, shoving the agentic capabilities of a capable LLM back into the chatbot style interface – even if that interface is more connected and capable than the web chat app – directs users back into simpler work.
Why don’t AI companies encourage sophisticated use?
The persistence of the chatbot-style interface for LLMs might suggest that the companies developing them aren’t interested in more sophisticated use. They’ve created something like a digital skeleton key that can theoretically operate any computer at the level of code: a powerful, capable technology that in the hands of creative users could be incredibly versatile. But when it comes to marketing this technology to the general public, it always returns to chatbots, answer engines, and writing marketing copy.
Why? If I’m being cynical, or perhaps just realistic, the majority of users probably couldn’t care less about an AI system that can write and execute its own code and “run tools in a loop”. Companies like OpenAI and Anthropic are measured on the number of active users, not the sophistication of use. OpenAI doesn’t seem to care that their customers are using AI for frankly terrible purposes like shopping, travel plans, and mental health advice, as long as they keep on using it.
The chatbot interface was a design accident – a fluke of OpenAI’s initial release of GPT-3.5 in November 2022. But it quickly drummed up a hundred million users, and that kind of success is hard to step away from. I suspect that, despite being a dead end, the chatbot interface will continue to haunt us for a long time yet.
As a result, the organisations that get disproportionate value from GenAI will be the ones that treat the interface as a starting point rather than the product, and that invest in the unglamorous skills Willison identifies: specifying goals, choosing tools, designing loops, and judging the quality of outputs. That last one, judgement, is the part that doesn’t come in the chatbot text box, and it’s where domain expertise becomes even more crucial.
Which brings me back to where the IYKYK series started. The premise in those articles was that the biggest barrier to AI capability was knowledge of what GenAI systems can actually do. Unfortunately, the interface is designed so you won’t ask those kinds of questions. If you want to really dig into the capabilities of modern AI systems, I’d encourage you to get out of the chatbot interface and into the slightly terrifying world of the command line. This is where the digital skeleton key starts to unlock some of the potential of these systems.
Speak to your local friendly IT nerd. They might be able to set you up with a safe-ish sandbox to play around in. You don’t need to become an expert in software design or command line interfaces, but I promise you that the experience of using a capable AI system in an environment other than the chatbot will open your eyes to the present and future capabilities of LLMs in a way that you’d never achieve with ChatGPT and Copilot.
Want to learn more about GenAI professional development and advisory services, or just have questions or comments? Get in touch:
Leave a Reply