Agentic CLI Coding Tools Converge on AGENTS.md and Agent Skills

The terminal became the main place AI coding agents compete

In late 2026 the command line is where AI coding agents compete hardest. An October roundup ranks Claude Code (running Opus 5.5), OpenAI's Codex (on GPT-6.1 Sol), the independent OpenCode project, Cursor CLI and Google's Antigravity CLI as the leading terminal agents. It also notes that DeepSeek's Harness, released August 13, has passed 247,000 GitHub stars in roughly two months.11 Another industry ranking calls the terminal "the new battleground" and argues that CLI agents get deeper system access, scriptability and CI/CD integration than tools that live only inside an IDE.13 A third comparison names Claude Code, Codex, Cursor, Copilot CLI, OpenCode and Antigravity CLI as the default shortlist for most teams.12

The number of tools matters less than what is happening underneath them. These agents now find their instructions and capabilities through shared file conventions instead of formats owned by a single vendor. Two standards do most of that work: AGENTS.md for project instructions and Agent Skills (SKILL.md) for reusable capabilities. Together with the Model Context Protocol (MCP), they are becoming a common discovery layer for every agent that runs in a shell. My reading is that this layer, more than any one model, is what will decide how portable agentic development ends up being.

Consolidation in the market makes portability more important

The market has been unstable this year. One analysis says 2026 brought heavy consolidation: SpaceX closed its acquisition of Cursor, and Google shut down Gemini CLI for consumers on June 18, 2026, replacing it with Antigravity CLI.11 Coverage does not fully agree here. A September ranking still lists Gemini CLI as the best free option, with a 1,000-request daily allowance.13 The most likely explanation is that tool rankings trail product changes, and that is a practical warning for teams choosing a tool.

When the tools change hands or shut down, developers need their project knowledge to stay put. The same October roundup tells developers to weigh who owns the vendor behind an agent and whether it supports MCP.11 Its top-ranked independent option, OpenCode, is MIT-licensed, supports more than 75 model providers and has over 212,000 stars.11 Another survey describes OpenCode similarly, with 208,000+ stars, 950 contributors and the ability to run offline.13 The numbers differ slightly, but both say the same thing: developers want tools that don't tie them to one vendor.

AGENTS.md: shared instructions for every agent

The first layer is project context. AGENTS.md is a plain Markdown file with no required fields, placed at a repository root. It tells agents how to build, test and modify a project.22 It was created to solve a specific problem: developers were keeping separate instruction files for each tool, such as .cursorrules, CLAUDE.md and Copilot's instruction file.22 The format came out of work involving OpenAI Codex, Amp, Google Jules, Cursor and Factory, and the Agentic AI Foundation under the Linux Foundation now maintains it.24

Adoption numbers are large but fuzzy. Several guides report that more than 60,000 open-source projects use the format.2429 One says 30-plus agents read it.22 Another notes that OpenAI's main repository alone contains 88 AGENTS.md files.24 In late 2025 the count was closer to 40,000 projects, so growth has been steady.28

The discovery rules themselves are fairly settled. Nested files are supported, and the file closest to the code being edited wins, which matters for monorepos.23 Codex reads a home-directory file and then walks the project tree, with a default combined cap of 32 KiB.24 Factory's Droid loads AGENTS.md at startup and also picks up nested files as it reads into subdirectories, with stated limits of 80,000 characters for the first load and 40,000 for dynamic discovery.25

The holdout: Claude Code and CLAUDE.md

The sources disagree most on one question: whether Claude Code, which several rankings call the consensus top agent, actually reads AGENTS.md. One guide says it does not do so natively. It reads only CLAUDE.md, and developers have to import or symlink AGENTS.md, although running /init will pull in an existing AGENTS.md.24 A compatibility table also marks Claude Code as not reading the file.29 Other write-ups describe support "via import"22 or as a fallback when no CLAUDE.md is present.26 One guide lists Claude Code among compatible runtimes without qualification.21

The documented workarounds suggest the cautious view is right. A standard that the most-used terminal agent supports only through a symlink is a widely adopted convention, not a universal one. Gemini CLI sits in between: it defaults to GEMINI.md, but a settings field can point it at AGENTS.md.2429 Community efforts have grown out of this gap. A proposed "Agents Standard" sets out a strict resolution order, including a global ~/.agents/AGENTS.md, an llms.txt product brief and a rule that agents never search above the project root. It ships a script that symlinks one file into the config locations of nearly twenty agents.27 A setup like that should not be necessary if the main standard were fully adopted.

Agent Skills: making capabilities discoverable

The second layer is capabilities. Anthropic launched Agent Skills on October 16, 2025, and published the format as an open standard at agentskills.io on December 18, 2025. That followed its donation of MCP to the Linux Foundation on December 9.4 A skill is a folder containing a SKILL.md file with YAML frontmatter, plus optional scripts and reference material.3

What matters for discovery is the loading model. At startup an agent reads only each skill's name and description, about 30–50 tokens per skill. The full instructions load only when a task matches, and reference files load only when needed.2 Clients also limit the listing itself. Claude Code gives it about 1% of the context window and caps each description at 1,536 characters. Codex keeps the list within 2% of the context window.3 As a result, a skill's description works almost like a search snippet competing for the model's attention. One guide puts it this way: the description "decides everything."3

Adoption has been fast, though the reported numbers vary. Accounts say OpenAI and Microsoft added support within about 48 hours, and Cursor followed in version 2.4 on January 22, 2026.4 One June guide counted 26+ adopting platforms2, while later coverage puts the client list at around 40 or more, including Codex, Copilot, Gemini CLI, JetBrains Junie, OpenCode, Kiro and Goose.345 Discovery is also becoming something agents do themselves. A "find-skills" skill lets an agent search for and install other skills when a user asks for something it can't yet do.6 Registries such as skills.md expose skills over both a CLI and MCP so agents can fetch them remotely when needed.10

Discovery paths still differ between agents

Shared formats do not mean shared locations. Claude Code looks in .claude/skills/. Codex scans .agents/skills/ from the working directory up to the repository root. Gemini CLI accepts either .gemini/skills/ or .agents/skills/. OpenCode checks all three conventions.7 The .agents/skills/ directory is becoming the neutral default, read by Codex, Cursor, Copilot, Gemini CLI and OpenCode, but Claude Code's documentation does not list it.7 One practitioner calls .agents/skills/ a universal alias that Claude Code can "theoretically" use.8 Given the documentation, that hedge is justified.

Frontmatter is another source of drift. Vendors add their own fields, such as Claude Code's subagent-forking and invocation controls or Codex's separate openai.yaml policy file, and those fields do not work in other agents.3 Only skills that stick to the six fields in the spec are fully portable.3 One analysis also warns that every installed skill becomes part of a developer's software supply chain.3 Because agents can now find and install skills on their own, that warning carries real weight.

What this means for developers

The coverage points to a stack with defined roles. MCP gives an agent access to external systems, skills supply reusable procedures, AGENTS.md describes how a repository works, and hooks or deny rules enforce hard limits the model cannot override.37 Shell-native agents are pushing that stack into automation too. Codex's codex exec, Kilo Code's --auto flag and Cline's headless mode all support running without a person at the keyboard.16

I don't think the important question in 2026 is which CLI agent wins. Ownership changes and shutdowns show that tools can disappear quickly. The more lasting result is that project knowledge and capabilities are moving into plain-text files that any agent can read. Interoperability is still incomplete, with differing directory paths, vendor-specific frontmatter and Claude Code's continued reliance on CLAUDE.md. Even so, the practical approach is to put instructions in AGENTS.md, keep skills in the spec's core fields, and treat each agent's own file as a pointer to that shared source.