The CLI agent wars settle into a pattern.

Two years ago, "agentic coding" was a research demo. One year ago, it was a feature war between editors. Now it has settled into something more interesting. Four well-resourced CLI agents (Claude Code, Codex, Aider, and Cursor's terminal mode) have independently arrived at nearly the same shape, and the differences between them have shrunk to questions of model preference and ergonomics. That convergence is the news.

When competing implementations from different teams converge, two things are usually true. The design space had a stable answer waiting to be found, and the next layer of innovation is going to happen above this one. Both are true here.

§01What converged

The shapes that all four CLI agents now share, more or less identically:

• Filesystem-scoped repo access. Each agent binds to a working directory and reads, edits, and searches files within that scope. None of them ask permission for trivial reads.
• Plan-then-execute discipline. Each one separates "decide what to do" from "do it." The naming differs (Plan mode, agent mode, /architect); the discipline is identical.
• Tool calls as the unit of work. Shell commands, web fetches, MCP servers; every agent expresses what it is doing as a sequence of tool calls a human can read.
• Conversational interface, scriptable underneath. Interactive by default, scriptable for automation.

The differences worth tracking:

• Model lock-in. Claude Code routes to Anthropic. Codex routes to OpenAI. Aider routes to anyone. Cursor routes to anyone, with a thumb on the scale for its proprietary endpoint.
• Open-source posture. Aider is fully open. Codex shipped open. Claude Code is closed. Cursor is closed.
• Ecosystem. Claude Code has the most mature skills, hooks, and MCP ecosystem. Codex is catching up fast. Aider has none, by design. Cursor lives in the editor, not the terminal-as-platform.

Strip the names off and the shapes are nearly the same.

§02Why convergence matters

Convergence is a signal, not a goal. It tells us the interface is settled. Once the interface stabilizes, energy moves up the stack to the things that use the interface: skills, hooks, statuslines, memory layers, evals, agent orchestrators.

That is why the mods directory and skills directory are growing faster than the agents themselves right now. The base layer (an LLM in a loop with tools) is a commodity. The differentiators are what gets wired on top of it: a team's prompts, the domain knowledge, the eval suites, the memory store.

§03What's coming next

Three predictions, in increasing order of certainty:

• Cross-agent skills become a standard. SKILL.md works in Claude Code today. The format is simple enough that other agents will adopt it (or a near-clone). Expect a 2026 to 2027 fork-then-converge cycle on a community standard.
• Memory becomes a first-class primitive. Today it is a community mod (Claude-Mem, dotfiles, a vector DB). Tomorrow it is agent.memory.read() and agent.memory.write() exposed by the runtime. Anthropic and OpenAI both have hints in this direction.
• Multi-agent orchestration becomes the new arms race. Once a single CLI agent is a commodity, the question becomes how five of them coordinate. Expect frameworks like LangGraph, Mastra, and CrewAI to absorb agent-coordination primitives that are currently DIY.

§04What it means for vibe coders

Practically, pick the CLI agent based on which model lab you trust most, and stop treating the choice as load-bearing. The deltas between Claude Code and Codex on the same task are smaller than the delta between two prompts you might write. The bigger choices are above the agent: which mods to install, which skills to encode, which evals to maintain.

The boring news is the news. The interesting period of CLI agents being different products is over. The period of them being a substrate for everything else is starting.