A tool gives an agent something it can do. A skill supplies a reusable way to approach work. A protocol defines how components communicate. Keeping these roles separate makes an agent easier to design and debug.
For an illustrative document agent, reading a file is a tool, an invoice review procedure is a skill, and MCP may connect the application to a document service. None of these establishes permission to disclose a document.
Tools are contracts, not names
A useful tool contract specifies its inputs, output shape, failure behavior and side effects. A schema can reject malformed arguments; it cannot establish that a valid request is appropriate for this user or task. Authorization must also be checked by the application or service.
As a proposed design checklist, make these distinctions explicit:
| Contract | Example | Why it matters |
|---|---|---|
| Scope | Search only this workspace | Prevents an ambiguous global search |
| Result | Identifiers, source locations and a continuation token | Lets later steps recover exact evidence |
| Failure | Denied, unavailable, empty or timed out | An empty result is not the same as a failed search |
| Side effect | Read, update or publish | Determines required controls and retry handling |
| Limit | Maximum results, bytes and runtime | Bounds work and context consumption |
Prefer explicit operations when a broad shell or general API would grant more authority than the task needs. Broad tools can be useful, but their flexibility also moves responsibility into execution controls. See Security & Control.
Skills package procedures
The Agent Skills specification packages metadata and instructions in a
SKILL.md file, with optional scripts, references and assets. Its progressive
disclosure pattern exposes identifying metadata first and loads detailed
instructions when relevant. A skill is not a new model weight or a security
boundary.1
For example, a review skill could explain which invoice fields to compare, where to record uncertainty and how to present evidence. A supporting script might implement exact arithmetic. The host still decides whether scripts may run and which files they may read.
Descriptions need clear applicability. Loading many overlapping procedures can introduce conflicting instructions and context overhead. Current OpenAI guidance recommends reconsidering excessive prompting and skill scaffolding as model capabilities improve; this is advice to test, not a guarantee that all shorter instructions are better.2
Protocols connect different boundaries
Figure 1 compares the endpoints rather than treating every protocol as a general agent integration layer.
- Agent applicationHosts a client
- Tool or context serverExposes supported capabilities
Connect a model application to tools, resources and prompts.
- Client agentRequests a task
- Independent agentRuns work and returns status or artifacts
Coordinate across agent implementations and services.
- Editor or UI clientDisplays progress and requests permission
- Coding agentRuns the session
Connect an editor interface with a coding agent. These standards do not share one interchangeable wire contract.
Figure 1. Protocol boundaries. Independent comparison of the official MCP, A2A and Agent Client Protocol descriptions.345
MCP standardizes access to tools and contextual information. It does not make a server trustworthy or decide whether an action is authorized. A2A supports communication with independent agents; it complements MCP rather than replacing it. ACP, here meaning Agent Client Protocol, connects coding agents to editor clients through session, progress and permission interactions.345
The July 28, 2026 MCP specification lists optional extensions for asynchronous tasks, skills and interactive applications. Support must be negotiated; do not assume that a client implements every extension or that an older tutorial matches the edition in use.3
Avoid making every tool result a model round trip
With ordinary tool calling, a model requests an operation and consumes its result before deciding again. Programmatic tool calling allows generated code to invoke tools, filter or aggregate intermediate data and return only the information needed for the next model decision. Tool discovery similarly loads relevant definitions on demand rather than inserting a large catalog into every prompt.6
An illustrative invoice check can compare thousands of line items in code and return the few unresolved matches. It need not place every record into the model’s context. The tradeoff is an additional code execution surface, which still needs authorization, resource limits and inspectable errors. Small tool sets and small results may not justify this extra machinery.
Choose the boundary before the technology
Ask whether the immediate need is a new capability, a repeatable procedure, a connection to a service or communication with another agent. Then choose the interface. Adding all four at once makes it harder to discover which change actually improved the system.
Continue with Context, Memory & State.
To see these ideas in a specific implementation, explore the Agent Deep Dives comparison and its five studies.
References
-
Agent Skills. Specification, file structure and progressive disclosure. Checked October 10, 2026. ↩
-
Eric Provencher, OpenAI. Rethinking skills and prompts for GPT-6 Astra, September 11, 2026. ↩
-
Model Context Protocol. Specification, July 28, 2026, core features and optional extensions. Checked October 10, 2026. ↩ ↩2 ↩3
-
A2A Protocol. A2A and MCP, complementary relationships. Checked October 10, 2026. ↩ ↩2
-
Agent Client Protocol. Protocol v1 overview. Checked October 10, 2026. ↩ ↩2
-
Anthropic. Introducing advanced tool use, 2025; tool search and programmatic calling. ↩