Technology is supposed to be boring. While our nerdy brains love to dive deep into technical implementations and architectural trade-offs, the general public does not.
So when technology attracts hype outside of tech circles, it’s a big deal. When the internet started to be mentioned in newspapers, we knew something was up. The same goes for the smartphone revolution and, yes, ChatGPT. A technology goes mainstream when the general public knows about it and uses it daily.
That is not the case yet for agentic AI, and at this rate, it is unclear whether it will ever become mainstream tech.
I first started hearing rumours about agents sometime in late 2023. Hypeboys on X would tout the coming singularity and the goal to build Artificial General Intelligence, the kind of AI that could take your job.
I remember being very sceptical about those claims because, while LLMs were good at generating text and mimicking conversations, I didn’t see their potential for tool-calling back then. It’s only when I saw an early version of Claude Code that I realised coding agents would change our industry forever.
But as amazing as coding agents are, we fail to see the similarly impactful equivalents in other fields. There are no successful autonomous agents in accounting, medicine or legal applications. That is: there are plenty of “ChatGPT for doctors”, but no Claude Code equivalent.
Part of that, without a doubt, is my lack of knowledge of those fields. But it’s also possible that the lightning that struck the software engineering field doesn’t strike twice.
Software development has hard feedback loops. It compiles, or it doesn’t. It has a culture of test automation and ubiquitous version control. This makes the impact of rogue agents small. If it goes off the rails, we just roll back the changes in GitHub.
Compare this to other fields. If a marketing agent makes a mistake, it might send out 100 emails to everyone in the entire database. People would mass unsubscribe, and there would be no undo button.
A legal agent could easily research and compile a contract. But the entire thing would have to be reviewed by a human, as there are no unit tests for legal docs.
Those industries handle that by adding human review and automated guardrails. Automated workflows with some agentic “tiny decision-making” sprinkled in. That is the right balance between quality and productivity. But it also ties the agent down. With these strict controls, it cannot get as creative as a coding agent.
Agents in software feel like senior engineers you can delegate work to. In other fields, they are intelligent workflow automation. The latter won’t take your job.
It unfortunately reminds me of the blockchain. Bitcoin and its clones were the one true application of this technology. Distributed payments were the only application that made sense. Every other idea, like NFTs and private blockchain ledgers, could be replaced by a simple database.
Those who saw the potential of the blockchain extrapolated that to a bunch of solutions that didn’t solve a real problem.
What if we are making that mistake again?
What if coding agents are the only real application for employee-like agentic AI?
What if coding was the only type of knowledge work that could be delegated completely to a machine?
I don’t have a crystal ball, and I’ve been wrong more than I’ve been right. But the longer we wait for AGI, the bigger the chance becomes that the machines will not take our jobs.
