Protocols & Coding Agents¶

Two adjacent topics: how agents talk to tools (and to each other) and what makes coding agents different. MCP was originally built for coding agents; coding agents use MCP heavily. Understanding either without the other is hollow.

Rapid Recall

MCP (Model Context Protocol) is "USB-C for AI", JSON-RPC 2.0 over stdio or Streamable HTTP, with three primitives (resources, tools, prompts). Solves the M×N integration problem; donated to Linux Foundation Dec 2025. A2A (Agent2Agent) is "HTTP for agents", task-stateful agent-to-agent protocol with signed agent cards at /.well-known/agent-card.json. The split: MCP for tools and data; A2A for agents. Coding agents have four 2026 paradigms: Autonomous cloud (Devin, fire-and-forget), Agent-first IDE (Cursor, Antigravity, parallel via git worktrees), Terminal/local (Claude Code, CLI + plan/build modes), Spec-driven (Augment, Intent, mandatory approval gates). Same brain, same tools underneath, different surfaces — pick by workflow.

The N+M insight¶

Before MCP, every team building serious AI tools was solving the same problem twice: once for OpenAI clients, once for Anthropic clients, once for whatever Cursor/Continue/Cline/local-models needed. Every integration was M × N — M hosts, N tools.

Without MCP (M × N integrations):           With MCP (M + N integrations):

   Claude  ─┬─ Notion                          Claude  ─┐
   ChatGPT ─┼─ GitHub                          ChatGPT ─┤
   Cursor  ─┼─ Stripe                          Cursor  ─┼─→ [MCP] ←─┬─ Notion server
   Custom  ─┴─ Linear                          Custom  ─┘           ├─ GitHub server
                                                                     ├─ Stripe server
   each host integrates each tool              host speaks MCP;      └─ Linear server
   M × N = 16 integrations                     each server speaks MCP
                                               M + N = 8 integrations

USB-C is the standard analogy. Before USB-C, every device had its own port; now any device that speaks USB-C plugs into any host that speaks USB-C. MCP is USB-C for AI.

flowchart TB
  subgraph BEF["Before MCP (M × N)"]
    H1[Claude] --- T1[Notion]
    H1 --- T2[GitHub]
    H2[ChatGPT] --- T1
    H2 --- T2
    H3[Cursor] --- T1
    H3 --- T2
  end
  subgraph AFT["After MCP (M + N)"]
    HH1[Claude] --> M[MCP layer]
    HH2[ChatGPT] --> M
    HH3[Cursor] --> M
    M --> S1[Notion server]
    M --> S2[GitHub server]
    M --> S3[Stripe server]
  end

Two protocols, two scopes¶

By early 2026, two protocols matter:

Protocol	Scope	Originated
MCP (Model Context Protocol)	Tools and data, how an LLM uses external capabilities	Anthropic, Nov 2024; donated to Linux Foundation Dec 2025
A2A (Agent2Agent)	Agent-to-agent, how an agent delegates to another agent across vendor boundaries	Google, Apr 2025; donated to Linux Foundation Jun 2025

The official line: MCP for tools, A2A for agents. Pick one based on whether the thing you're calling is a passive tool (it does what you ask, returns a result) or an active agent (it has its own loop, its own tools, makes its own decisions, can take time, can stream updates).

Section guide¶

Page	Covers
MCP	Architecture (host/client/server), three primitives (resources/tools/prompts), transports (stdio + Streamable HTTP), OAuth 2.1 auth, minimal FastMCP server, security model
A2A	Why a second protocol, agent cards, signed discovery, task lifecycle, composition with MCP
Coding Agents	What makes coding agents structurally different, the four paradigms (Devin / Cursor+Antigravity / Claude Code / Augment), the canonical toolkit, context management, evaluation

Protocols & Coding Agents¶

The N+M insight¶

Two protocols, two scopes¶

Section guide¶

Layer Checklist¶