# AI agent terms explained: MCP, A2A, skills, memory and evals

Understand the protocols and parts of an AI agent, with practical examples, primary sources and questions to ask before choosing a tool.

By Agent Search Engine. Sources checked 2026-10-28. AI-assisted editorial explanation; examples below are illustrative, not hands-on vendor tests.

## Which problem does each protocol solve?

- **MCP:** An AI application to tools and context. Illustrative example: Read issues through a GitHub connector. [MCP specification release · 28 July 2026](https://blog.modelcontextprotocol.io/posts/2026-07-28/)

- **A2A:** Independent agents to one another. Illustrative example: Delegate a task to another agent. [A2A specification](https://a2a-protocol.org/latest/specification/)

- **Agent Skills:** An agent to reusable task instructions. Illustrative example: Load a team's release-review procedure. [Agent Skills specification](https://agentskills.io/specification)

- **WebMCP:** An agent to actions exposed by a website. Illustrative example: Use a site's structured search action. [WebMCP draft · 27 October 2026](https://webmachinelearning.github.io/webmcp/)

## What is MCP?

Model Context Protocol lets AI applications connect to tools and contextual resources exposed by servers. A client chooses what it supports and how permissions work. The published 28 July 2026 specification changes the request model and develops extensions and authorization; support must still be checked in each implementation.

**Illustrative example:** A coding assistant reads an issue through an MCP server, then proposes a patch. The connector supplies access; it does not prove the patch is correct.

**Ask before choosing:** Which client, server and protocol version were tested? What can the connection read or change?

Sources: [MCP specification release · 28 July 2026](https://blog.modelcontextprotocol.io/posts/2026-07-28/)

[Browse MCP servers](https://agentsearchengine.app/category/mcp-servers); [Connect an MCP server](https://agentsearchengine.app/guides/connect-mcp-server)

## What is A2A, and how is it different from MCP?

Agent2Agent is a protocol for agents to discover one another and exchange tasks, messages and artifacts. Its Agent Card describes a provider's capabilities and connection details. MCP gives an application tools and context; A2A gives independent agents a way to communicate. Neither guarantees that a delegated task succeeds.

**Illustrative example:** A planning agent asks a separate research agent for a report. That researcher might use MCP tools while doing the work, then return an artifact through A2A.

**Ask before choosing:** Is communication authenticated, and can you inspect what the other agent did before accepting its result?

Sources: [A2A specification](https://a2a-protocol.org/latest/specification/)

[Browse agent frameworks](https://agentsearchengine.app/category/agent-frameworks)

## What are Agent Skills?

Agent Skills packages reusable instructions with optional references, assets and scripts. The open format uses a SKILL.md file. Clients can load a short description first and detailed instructions when needed, a pattern called progressive disclosure. A skill can tell an agent how to use an MCP tool.

**Illustrative example:** A release-review skill supplies your checklist and report format; an MCP connection supplies the repository information needed to follow it.

**Ask before choosing:** Who maintains the skill, does it include executable scripts, and which permissions or dependencies does it require?

Sources: [Agent Skills specification](https://agentskills.io/specification); [Agent Skills overview](https://agentskills.io/home)

[Explore coding agents](https://agentsearchengine.app/category/coding-agents)

## What is WebMCP?

WebMCP lets a website expose structured actions to visiting agents. Chrome describes declarative form integration and imperative JavaScript APIs. This is different from an external MCP server or clicking a screenshot. The 27 October 2026 Community Group draft is not a W3C Standard; implementations are still developing.

**Illustrative example:** A travel site could expose a search action with named date and destination fields. This is an illustrative design, not a claim about a particular live site.

**Ask before choosing:** Which browser and agent support it today? What happens if the action is unavailable or the visitor declines it?

Sources: [WebMCP draft · 27 October 2026](https://webmachinelearning.github.io/webmcp/); [Chrome WebMCP early preview](https://developer.chrome.com/blog/webmcp-epp)

[Browse browser and computer-use agents](https://agentsearchengine.app/category/browser-agents)

## What is an agent harness?

The harness is the surrounding software that gives a model tools, context, an execution environment and a way to continue working. It helps turn model output into a working process. Harness design includes feedback and recovery; a model's benchmark result does not describe the whole system.

**Illustrative example:** A coding harness can edit files, run checks, read the failure and attempt a correction. The same model in a simple chat box does not have that entire workflow.

**Ask before choosing:** Can the agent recover from failures, stop at a limit, and show evidence of what changed?

Sources: [OpenAI: harness engineering](https://openai.com/index/harness-engineering/)

[Choose a coding agent stack](https://agentsearchengine.app/guides/coding-agent-stack)

## What does context engineering mean?

Context engineering means selecting and maintaining the instructions, tools, history and retrieved information available to a model. It goes beyond wording one prompt. The context window is finite: too much irrelevant or stale material can make the task harder. The cited foundational article was published in 2025, not invented this year.

**Illustrative example:** For a support reply, provide the current refund policy and the relevant order facts, rather than an entire document archive and every past conversation.

**Ask before choosing:** Where does the information come from, when was it updated, and is this user allowed to access it?

Sources: [Anthropic: effective context engineering](https://www.anthropic.com/research/effective-context-engineering-for-ai-agents)

[Build a knowledge assistant](https://agentsearchengine.app/guides/rag-stack)

## What is RAG, and what is agentic retrieval?

Retrieval-augmented generation supplies retrieved information to a model when it answers. Retrieval can use document search, a database or other sources; it is not synonymous with a vector database. With agentic retrieval, the agent decides when to search, what to request and whether another search is needed.

**Illustrative example:** A research agent finds an announcement, notices a missing pricing detail and performs another search before answering. More searches alone do not make the answer better.

**Ask before choosing:** Can you trace the answer to the retrieved evidence and reproduce a failure on a fixed test case?

Sources: [LangChain: context and retrieval · 23 October 2026](https://www.langchain.com/blog/context-engineering-starts-with-retrieval)

[Explore research agents](https://agentsearchengine.app/category/research-agents); [Read the RAG stack guide](https://agentsearchengine.app/guides/rag-stack)

## Is agent memory the same as context or RAG?

Memory is information retained for later use, such as thread state or facts stored across sessions. Context is what the model can see for its current work. Retrieval can bring stored memories into that context. Saving a fact does not itself update the model's weights or guarantee accurate recall.

**Illustrative example:** An assistant stores a preference for concise reports, then retrieves it for a later session. A correction or deletion needs to reach the stored record too.

**Ask before choosing:** What persists, who can read it, can you correct or delete it, and does it cross user or workspace boundaries?

Sources: [LangChain: memory concepts](https://github.com/langchain-ai/docs/blob/main/src/oss/concepts/memory.mdx)

[Browse agent infrastructure](https://agentsearchengine.app/category/agent-infrastructure)

## What are agent evaluations, or evals?

Evals check whether an agent behaves as intended on representative tasks. Offline evaluation compares versions on prepared cases; online evaluation examines production behavior. Test the actions and final environment state when they matter, not just a plausible final answer. A published benchmark cannot settle performance on your own workflow.

**Illustrative example:** A booking test checks the saved date, budget and approval record. A confident message saying the booking succeeded is insufficient if the underlying record is wrong.

**Ask before choosing:** Which real tasks and failure cases were tested, and what did the checks actually measure?

Sources: [LangChain: evaluation types](https://docs.langchain.com/langsmith/evaluation-types); [LangChain: evaluating agent runs · 15 October 2026](https://www.langchain.com/blog/common-questions-about-agent-evals)

[Browse agent infrastructure](https://agentsearchengine.app/category/agent-infrastructure)

## How is observability different from evaluation?

Observability records what happened during a run, such as model requests, tool calls and errors. Evaluation judges whether the behavior met the task's requirements. A trace can help explain a failure without proving quality. OpenTelemetry's GenAI conventions remain active work; inspect the current version before relying on particular fields.

**Illustrative example:** A trace reveals that a tool timed out. An evaluation shows whether the agent handled that timeout correctly and avoided falsely claiming success.

**Ask before choosing:** Can you inspect useful traces without retaining unnecessary personal information or secrets?

Sources: [OpenTelemetry: GenAI semantic conventions](https://github.com/open-telemetry/semantic-conventions-genai)

[Browse agent infrastructure](https://agentsearchengine.app/category/agent-infrastructure)

## What should agent guardrails actually protect?

Guardrails are controls that limit or check access and actions. OWASP's agentic risk list covers issues including tool misuse, privilege problems and memory poisoning. An agent can produce a useful answer while taking an unauthorized action, so permissions, validation and consequential-action approval need their own checks.

**Illustrative example:** An assistant drafts a refund but needs an authorized person to approve issuing it. A free-text promise to be careful is weaker evidence than an enforced permission boundary.

**Ask before choosing:** What can the agent change, which actions require confirmation, and how can you stop or recover a failed run?

Sources: [OWASP: agentic application risks, 2026 edition](https://genai.owasp.org/resource/owasp-top-10-for-agentic-applications-for-2026/)

[Read about autonomous agent stacks](https://agentsearchengine.app/guides/autonomous-agent-stack)

## What is a computer-use agent?

A computer-use agent interacts with application interfaces using information such as screenshots, the DOM or accessibility tree. It differs from calling a purpose-built API. Clear controls and stable layouts help interaction, but interface access alone does not establish reliable task completion or appropriate permissions.

**Illustrative example:** An assistant navigates a web form that has no integration. It still needs to identify the correct fields, handle errors and stop before a consequential submission when approval is required.

**Ask before choosing:** Does the complete workflow survive an error, a changed layout and the required approval step?

Sources: [web.dev: agent-friendly site UX](https://web.dev/articles/ai-agent-site-ux)

[Choose a browser-automation stack](https://agentsearchengine.app/guides/browser-automation-stack)

## Two different protocols called ACP

ACP has two distinct meanings here. Agent Client Protocol connects coding agents and editors; its docs say full remote support remains work in progress. Agentic Commerce Protocol concerns commerce between agents, buyers and businesses. Neither name implies compatibility with the other, and protocol implementation alone does not secure platform distribution.

Sources: [Agent Client Protocol introduction](https://agentclientprotocol.com/get-started/introduction); [Agentic Commerce Protocol](https://www.agenticcommerce.dev/)

## What about world models and other new announcements?

World models can represent or simulate environments. DeepMind's 27 October 2026 Genie 4 announcement describes interactive world simulation as a trusted-tester research preview. Treat a research preview, a public beta and a generally available product as different stages; an announcement does not establish availability or reliability for your task.

Source: [DeepMind: Genie 4 research preview · 27 October 2026](https://deepmind.google/blog/genie-4-a-new-generation-of-world-simulation/)

## Turn terminology into a useful choice

Editorial checklist; it is not a vendor certification.

1. Define one task, its input and what a successful result would change or produce.

2. Check the exact product, client, version and release stage against primary documentation.

3. Inspect permissions, data destinations, retention and approval for consequential actions.

4. Try a representative case, an error case and an ambiguous request in an isolated environment.

5. Record actual outputs, actions, limits and total operating costs before expanding use.

Source descriptions establish what the provider or specification says, not independent performance, adoption or safety. Release stages and compatibility can change after the checked date.

Editorial method: https://agentsearchengine.app/methodology

Source: https://agentsearchengine.app/guides/ai-agent-terms