Deep Agents Explained: Skills Manage Internals, MCP/A2A/ACP Manage Externals

Deep Agents consist of four distinct mechanisms that operate on different dimensions: Skills (an internal specification), MCP (tool‑integration protocol), A2A (agent‑to‑agent collaboration protocol), and ACP (client‑to‑agent exposure protocol).

Key Distinction

Skills is the only "inward" mechanism; it does not connect to external systems but defines the standards an agent follows internally. MCP, A2A, and ACP are "outward" protocols that manage connections to tools, other agents, and clients respectively.

Why Skills Is Separate

Skills is not a protocol but an operation manual. Each Skill is represented by a SKILL.md file that lists processes, checklists, and standards. The agent decides when to trigger a Skill based on a description field and then follows the prescribed steps. Unlike MCP, which provides the ability to execute SQL, a Skill might require the agent to check for SQL‑injection risk before execution.

Four‑Layer Interaction Diagram

┌─────────────────────────────────────┐
│  Client (IDE / CLI / Web)           │
│          ↕ ACP                      │
│  Agent Service                      │
│   ┌─────────────────┐              │
│   │  Skills (internal) │          │
│   └─────────────────┘              │
│    ↕ A2A      ↕ MCP                 │
│  Other Agents   Tools / Resources   │
└─────────────────────────────────────┘

The diagram shows that Skills sits inside the agent, while the three protocols form connections outward.

Step‑by‑Step Scenario: Enterprise Code Review

User initiates request : The IDE calls the review agent via ACP .

Agent follows a Skill : It loads the code-review Skill, which defines a checklist (bug correctness → boundary conditions → test coverage → security risk). The agent executes the steps in order.

Security scan needed : When reaching the security step, the agent uses A2A to locate a dedicated security‑scan agent and forwards the code.

Run scan tool : The security‑scan agent connects to a SAST tool via MCP , performs static analysis, and returns a vulnerability report.

Return results : The report travels back along the same path: security agent → review agent → IDE.

When to Use Each Layer

Teams typically do not need all four layers at once. Use the following guide based on current pain points:

If agent output quality is unstable, focus on Skills .

If the agent needs to read databases, call APIs, or access files, adopt MCP .

If multiple agents must cooperate, enable A2A .

If IDE/CLI/Web clients need a unified entry point to the agent, implement ACP .

Prioritization Recommendation

In practice, start with Skills plus MCP . Adding a SKILL.md file immediately stabilizes agent output, and MCP provides mature tool integration covering about 80 % of scenarios. Add A2A and ACP later when multi‑agent collaboration or multiple client entry points become necessary.

Author’s Final Insight

The most important takeaway is to separate internal quality concerns (Skills) from external connectivity concerns (MCP, A2A, ACP). Treat them as distinct layers rather than competing priorities.