Phodal Huang, who introduces himself in the post as the author of the open-source AI coding tool AutoDev, frames this year-end review around a case study he had read of a Chinese AI coding tool that spent considerable space on the agent’s “added value” of comforting and encouraging a programmer who is struggling. His response is a question: if the models and tools were really strong enough, why would the experience need to be made up for with comfort? What a programmer needs when a problem will not resolve, he argues, is a system that gets the job done.
The post then sets out six trends he saw in AI coding in 2025, with an emphasis on the Chinese ecosystem, and closes by saying that in 2026 developers do not need a chat box that says “keep going” but a “digital partner” that finds a system flaw at two in the morning and quietly submits a fix — and that AutoDev’s own path is to give the bandwidth spent on “emotional companionship” back to “engineering certainty”.
Key Points
- Chinese coding models have begun following Claude’s route of strengthening the coding ability and agentic behaviour of their main text models. Having tried MiniMax M2.1 and GLM 4.7 at year end, the author reports solid agentic behaviour — clear task planning, sensible tool calls, multi-step engineering operations — while noting a legacy-system migration scenario in which the model lost focus after a long conversation, and cautioning that high leaderboard scores do not carry over to every scenario.
- Many model vendors have gradually given up building their own AI coding tools and instead make their models usable inside mature existing tools, so that developers get the most mature tool plus a replaceable model — for example Chinese models inside Cursor or Claude Code.
- Specifications are making a comeback alongside continued experimentation with Context engineering: he groups the rise of Model Context Protocol and then of Skills, AGENTS.md and Spec-driven development as coding tools exploring what better context engineering looks like. He points to Claude Code adding LSP support for gathering context at the end of the year, and calls file-operation-based agentic RAG too costly and prone to missing key context.
- The barrier to end-to-end tooling is falling: he cites Atlassian’s Rovo Dev CLI and its integration with Jira and Bitbucket, GitHub Copilot’s deeper integration with GitHub.com, Augment Code’s review product, and Cursor’s December acquisition of the code-review company Graphite, concluding that competition is moving from single capabilities to end-to-end closed loops.
- In what he calls an era of self-verification, tools have begun to take responsibility for their output — judging whether a task is actually complete rather than only whether the code runs. He cites testing agents such as Playwright’s native agents and ScenGen using OODA-like loops, and Playwright’s Healer Agent replaying failed steps to generate fixes.
- AI lowers the barrier to every technology stack, which he says also means AI is replacing repetitive, templatable work; the valuable work that remains is designing complex systems, planning engineering delivery and coordinating upstream and downstream processes.
Context
The post is an opinion piece grounded in the author’s own trials of tools and models; it reports no measurements. On code review he describes having shown Augment’s team, through enterprise access they provided, what he thought an AI-era code-review agent should do — a structured summary of the whole change, a diagram-level overview rather than fragmented comments, and continuity across the change’s lifecycle — and notes that the product now implements this, without claiming the direction came from his feedback.