Episode 2: Common Ground - Shared Capabilities and Limitations

Despite their different form factors, by 2026, the industry has reached a high degree of consensus on the underlying logic of AI Coding Agents. Regardless of which one you choose, the following core capabilities are "standard."

✅ Core Capabilities Supported by All

1. Sub-agent / Multi-Agent Collaboration: Support for delegating complex tasks to sub-agents that execute independently and return results.
2. MCP (Model Context Protocol): The de facto standard for unifying access to external tools (e.g., MySQL, Jira, search engines).
3. Skills Concept: Support for reusable atomic capabilities defined based on the Agent Skills Open Standard.
4. Slash Commands: Quick triggering of preset workflows by entering commands like /fix or /test.
5. Cross-Tool AGENTS.md Standard: Support for using a single Markdown file to issue unified project rules to all three tools.
6. Multimodal Input: Understands not just code, but also images (design drafts) and PDF documents.
7. Pause and Approval Flows: Support for requesting user approval before executing high-risk operations (like writing files or running Bash).
8. Markdown + YAML Extension Pattern: Uses the formats most familiar to developers to define rules and workflows.

⚠️ Shared Limitations

Understanding the boundaries of capability is equally important. Currently, all three tools share the following limitations:

• Primary Agent is Non-Replaceable: You cannot completely replace the product's built-in "main brain" with your own logic; all extensions are built on top of the official primary Agent.
• System Prompt Black Box: The underlying System Prompt that actually controls Agent behavior is controlled by the vendor; users can only perform "instruction injection" through rule files (CLAUDE.md, GEMINI.md, AGENTS.md).

This consensus means: If you master extension development for one (such as writing a Skill or Rule), the barrier to migrating to the other two will be very low.