The year the coding agent grew up — and started breaking things
Agentic coding tools went from autocomplete curiosities to background workers with production access. The productivity wins are real. So are the outages.
Something shifted in how developers actually work this year, and it's worth naming plainly. Simon Willison captured it in a throwaway post this week: he's shipping better code now because tedious 20-file readability tweaks are suddenly free. "Costs me nothing," he wrote. That phrase — costs me nothing — is the quiet revolution. When the marginal cost of a minor improvement drops to zero, the bar for what counts as "worth doing" collapses, and codebases start getting the kind of care they never used to.
The stack is consolidating, the stakes are rising
The tooling landscape has sorted itself out. A dev.to roundup this week calls Cursor the best-in-class IDE-native option while noting it's "a fast, well-designed product built on top of other people's models" — a reminder that the interesting moats now live in workflow, not weights. Underneath, the model layer is crowded: OpenAI just shipped GPT-5-codex, a variant optimized for agentic coding in Codex, defaulting it for cloud tasks and code review. Anthropic continues pushing Claude Code as a terminal-native agent that handles git workflows through natural language. Alibaba released Qwen3.6-Plus with stronger agentic capabilities, and Google DeepMind shipped models in four sizes — including 31B dense and 26B MoE variants aimed squarely at local coding assistants.
But the same dev.to piece draws a line that deserves more attention than it's getting: "Agents with write access to production databases have destroyed real systems in 2026." The cost of an AI coding mistake is no longer a suggestion you ignore — it's a background agent that finishes a task you'd have stopped if you'd been watching.
What this means depends on where you sit
For builders, the productivity delta is genuine, but the discipline question has inverted. The old worry was whether the AI would write good code. The new worry is whether you're reviewing what it did before it touches something that matters. For engineering leaders, the governance conversation is overdue: which agents have write access to what, and who signs off on autonomy boundaries? For everyone else — the product managers, the designers, the operators watching this from the sidelines — the trajectory is that fully autonomous agents taking a Jira ticket to a merged PR will be standard by late 2026. That changes what "engineering capacity" means as a planning input.
The bottom line: Treat every coding agent like a new hire with root access — the productivity is worth it, but only if you've decided in advance what it's allowed to touch unsupervised.