Skip to main content
What this unit solvesWith too many tools and updates moving too fast, instead of memorizing product names, build a map that classifies tools by interface type and lets you place any new tool quickly. This unit gives you that map and marks the main product matrix as of 2026-05 along with the cross-tool standards that are converging.Note: everyone has their own way of placing tools, so the ability to make independent judgments is essential.

​
Learning objectives

  • Classify any AI tool by “interface type” (conversational assistant / CLI agent / IDE agent / desktop agent / cloud agent).
  • Sketch the product matrix for the four major vendors plus Cursor, and describe each one’s positioning.
  • Explain why the AGENTS.md cross-tool standard matters and what it is converging toward.
  • Distinguish the cost structures of the “subscription” and “API billing” usage models.

​
1. Classification by interface type (the core map)

Memorizing product names is futile; this field ships something new every two weeks. What is genuinely stable is interface type: the form through which a model interacts with you, what it can touch, and where you intervene. Build this map first, and when any new tool arrives you only need to ask “which type is it?” to know roughly how it is configured, where the risk lies, and how it connects to your existing workflow. Two criteria draw the lines: degree of human intervention (do you approve every step, or dispatch a task and walk away?) and action scope (text-only output, or can it read and write your files, run commands, open PRs?).
TypeHuman interventionAction scopeRepresentatives (as of 2026-05)
Conversational assistantWatching every stepText output onlyClaude.ai, ChatGPT, Gemini
CLI agentApproval checkpointsRead/write files, run commandsClaude Code, Codex CLI, Antigravity agy CLI
IDE agentIn-editor approvalWorkspace filesCursor, Antigravity, Copilot
Desktop agentGoal-directed, minimal interventionLocal files and applicationsClaude Cowork
Cloud agentAsync after task dispatchCloud sandbox, open PRsCopilot coding agent, Codex cloud, Jules
Type determines the risk surface, not the brandA single vendor spans multiple types (OpenAI has conversational ChatGPT, CLI-based Codex, and cloud-based Codex cloud). What you should watch is action scope: a conversational assistant can at most give you wrong information, while CLI and cloud agents can modify your files, run your commands, and touch your repo. The wider the action scope, the less you can skip on upfront configuration (permissions, sandboxing, approval strategy). This line runs through all of Part II and 03-3.

​
2. Product matrix of the four major vendors (as of 2026-05)

Map the five types onto vendors and you get this matrix. The point is not to memorize it, but to see the trend: every vendor is expanding into multiple types, and none stays in just one.
Interface typeAnthropic Claude (primary reference)OpenAIGoogleGitHub / MicrosoftCursor (Anysphere)
Conversational assistantClaude.aiChatGPTGeminiCopilot Chat(built-in chat)
CLI agentClaude CodeCodex CLIAntigravity agy CLICopilot CLI(no standalone CLI)
IDE agent(via Claude Code / extension)Codex IDE extensionAntigravity IDECopilot (VS Code)Cursor itself
Desktop agentClaude Cowork(no equivalent standalone product)Antigravity desktop app(no equivalent standalone product)(none)
Cloud agent(Claude Code cloud execution)Codex cloudJulesCopilot coding agentBackground agents
The flagship model version column has the shortest shelf life; don’t memorize itModel designations change faster than products do. As of 2026-05, Anthropic’s flagship is the Claude 4.x family (Opus 4.8 / Sonnet 4.6 / Haiku 4.5), and Google Antigravity 2.0 is built on Gemini 3.5 Flash. For OpenAI GPT generation versions, consult the official changelog. When evaluating a tool, look at interface type and configuration mechanisms; treat model versions as floating parameters to look up when needed.
Two naming traps to clear up first (carried over from the README Section 5 clarification):
  • Claude Cowork is not Claude Code. Cowork is a desktop agentic product for knowledge workers (goal-directed, touches local files and applications); Code is a CLI for developers. Different products, different interface types.
  • Cursor is a third party. Developed by Anysphere, not a first-party product of any model vendor; it connects to multiple vendors’ models.

​
3. AGENTS.md: the rise of a cross-tool standard

In the past, every tool had its own project-level rules file: Claude’s CLAUDE.md, Gemini’s GEMINI.md, Cursor’s .cursor/rules/, Copilot’s .github/copilot-instructions.md, and OpenAI Codex’s AGENTS.md. Change tools and you had to rewrite all your rules. That fragmentation is converging, and it is converging toward the AGENTS.md format that Codex pioneered. AGENTS.md is an open Markdown format for giving agents project-level instructions and context (coding conventions, build steps, testing requirements). It originated in OpenAI Codex’s toolchain, released in 2025-08, and by 2026 had been adopted by 60,000+ open-source projects and frameworks, including Codex, Cursor, GitHub Copilot, Jules, VS Code, Gemini CLI, Amp, Devin, Factory, and others [1]. Governance is the key to its viability as a standard: in 2025-12, AGENTS.md was donated to the Linux Foundation alongside MCP and goose, forming the newly created Agentic AI Foundation (AAIF), co-founded by OpenAI, Anthropic, and Block, with support from Google, Microsoft, AWS, Bloomberg, Cloudflare, and others [1]. In other words, several competing vendors handed stewardship of the project-rules-file standard to a neutral foundation. That is the signal that it will not be captured by any single vendor.
What this means in practiceThe trend is that project-level rules are converging from “one filename per vendor” toward a single AGENTS.md, lowering cross-tool migration costs. But convergence is not complete: each tool’s priority ordering between AGENTS.md and its own proprietary file still differs (Antigravity treats GEMINI.md as an override/supplement to AGENTS.md without declaring a strict priority order; Copilot reads both AGENTS.md and CLAUDE.md/GEMINI.md, with non-deterministic conflict resolution; Cursor treats AGENTS.md as a simplified substitute for .mdc). The pragmatic approach now: write AGENTS.md first for shared rules on new projects, and put the small vendor-specific differences in each vendor’s proprietary file. For precise priority ordering and filenames per tool, see 02-6.

​
4. Subscription vs. API billing

The same model reached through different billing structures has completely different cost profiles. Getting this wrong will trip you up the moment you want to automate something.
  • Subscription (ChatGPT Plus/Pro, Claude Pro/Max, Gemini AI Pro/Ultra, etc.): flat monthly fee, with usage limits that are usually rolling time windows (e.g., messages or token quota per 5 hours or per week). Suited to interactive, human-in-the-loop daily use. Hit the cap and you wait for the window to reset.
  • API billing: charged per token (input and output priced separately), no interactive caps, suited to automation, batch processing, and high-volume programmatic calls. An agent loop running all night belongs on API billing, not forced through a subscription seat.
A third path for researchers: local inferenceBeyond subscription and API, running open-weight models on your own hardware (vLLM, SGLang, llama.cpp) is a third option. The cost is operating and tuning the stack yourself; what you gain is two things: cost (one-time hardware outlay replaces an ongoing token bill, which can be more economical at high inference volumes) and privacy (sensitive data never leaves your machine and never passes through a third party). For research workflows handling unpublished data or regulated datasets, the privacy property is often the deciding factor, not the cost. Which of the three paths to choose depends on your usage curve and data sensitivity, not on which one is “more advanced”.

​
5. How to keep tracking the landscape (without chasing stars)

The signal-to-noise ratio in this field is extremely low. Second-hand “annual roundups” and “strongest tool rankings” are mostly SEO content farms or marketing reach, not inputs for your decision-making. Build your own low-noise channel:
  • Follow official primary sources: each vendor’s changelog and release notes, official engineering blogs, YouTube. Primary information is more accurate and faster than any second-hand summary.
  • Track authoritative or influential individuals: such as known core engineers at Anthropic, and others you find credible.
  • Filter with the classification map: when you see a new product, first ask “which type is it?” not “how strong is it?” Once you have placed it, you know whether to look further. Most new releases are irrelevant to you; classification lets you screen them out in three seconds.
  • Hand “is it worth investing in” to a framework: whether a tool is worth learning does not depend on star count or others’ recommendations, but on whether it genuinely helps with your tasks. You can also extract ideas from high-star or recommended projects and adapt them to something that fits you better. That judgment framework is at 03-2.
Placing an unfamiliar tool with the map: agyGoogle I/O in 2026-05 released Antigravity 2.0, and one new item in it is called agy. Suppose you only see the name; run it through the map:
  1. Which type? It is a CLI (invoked from the command line, rewritten in Go for low memory and fast startup), so it belongs in the CLI agent row, with action scope of reading/writing files and running commands.
  2. What does it replace? It is the successor to Gemini CLI, which reached end-of-service for AI Pro/Ultra on 2026-06-18 [2]. If you were using Gemini CLI, this is your migration target.
  3. What does the configuration roughly look like? The shared configuration surface for CLI agents (rules files, permissions/sandboxing, MCP) is covered in Part II; the specific filenames for the Antigravity system (~/.gemini/GEMINI.md, workspace .agents/, MCP config ~/.gemini/config/mcp_config.json) are in 02-6.
  4. Is it worth investing in? Depends on whether you are in the Gemini ecosystem and whether you need its parallel sub-agents. Use the 03-2 judgment framework.
After the full placement, you have used a stable map to put a changing product into known coordinates. That is the entire point of this unit.

​
Common pitfalls

Anti-pattern list
  • Memorizing tools by vendor: “the OpenAI one”, “the Google one” is impossible to track and maps poorly, because every vendor spans multiple types. Memorize by interface type and the map stays stable.
  • Treating a benchmark ranking as “better for my task”: leaderboards measure average tasks, not your tasks. A model at the top of the leaderboard does not mean it is more useful in your domain or workflow. See 03-2 for the judgment approach.
  • Chasing every new release: every new tool carries migration cost and learning cost. When you see something new, place it with the map first, then ask “does it solve an existing pain point of mine?” If not, skip it. FOMO is not a reason for tool selection.
  • Assuming AGENTS.md is already a unified standard: convergence is in progress but not complete, and priority ordering differs across tools. Use it as a “shared layer”, but don’t assume behavior is consistent everywhere.

​
Self-check

The bar for passing this unit
  1. Given a tool name you have never used, can you say within 30 seconds which type it belongs to, how wide its action scope is, and roughly what its configuration looks like?
  2. Can you explain the difference between Claude Cowork and Claude Code, and why Cursor doesn’t count as a first-party product of any vendor?
  3. Can you explain what AGENTS.md is converging, who governs it, and why that matters?
  4. For an automation task that needs to run all night, would you use a subscription seat or API billing? Why?

​
Sources and further reading

Fast-moving facts are marked as of 2026-05; model versions and product details should be verified against official changelogs.