LLM Predictions for 2026

Simon Willison predicts that 2026 will be the year coding agents become undeniable. He suggests we are currently running these agents with far too much access to our systems, effectively operating on a "it hasn't broken yet" security model. He expects a major security incident to force the industry to finally solve the problem of sandboxing and isolation for untrusted code.

"I think we're due a Challenger disaster with respect to coding agent security... so many people, myself included, are running these coding agents practically as root."

Looking further ahead to 2029, Willison points to the Jevons paradox as the defining economic question. If AI makes it ten times cheaper to produce code, the demand for software may simply increase tenfold. If this happens, the role of the engineer shifts from writing lines of code to verifying the result and securing the boundaries. It suggests that our value in the future will be found in how we defend the constraints of a system, rather than how fast we can type.

I have explored similar themes regarding the shift from manual checking to structural safety in Guardrails over Gates and the changing role of the engineer in The New Mandate.