Most of the systems an AI agent needs to reach were built before anyone planned for agents. The order history sits in an ERP. Eligibility rules live in a database that three applications share. Nobody is going to rewrite those systems this year, and nobody needs to.
O'Reilly's Signals for 2026 says architects are looking for ways to modernize existing systems for AI, and that the rise of agentic AI means keeping up with emerging protocols like MCP (O'Reilly). I agree with the direction. The work is less about new technology than about giving agents a controlled way into what you already run.
What MCP is
The Model Context Protocol is an open standard for connecting AI applications to tools and data. Anthropic introduced it in November 2024. An MCP server describes each tool it offers with a name, a plain-language description, and an input schema. An agent reads that description and decides when to call the tool.
The reason to pay attention is adoption. In December 2025 Anthropic donated MCP to the Agentic AI Foundation, a new fund under the Linux Foundation, with Amazon Web Services, Google, Microsoft and OpenAI among the members. The Linux Foundation counted more than 10,000 published MCP servers at the time, and said Claude, ChatGPT, Gemini, Microsoft Copilot and VS Code had all adopted the protocol (Linux Foundation). An integration you build once can now serve several AI clients.
Inventory what an agent would need to touch
Start with a list of systems. For each one, write down four things: what an agent would read from it, what an agent would change in it, how sensitive the data is, and whether a documented API already exists.
That list sorts itself. A system with a documented API and low-risk data is a first candidate. A system you can only reach through its database or its screens needs an API in front of it before any agent gets near it. That API is modernization work you would want anyway.
Wrap what you have before you rewrite
On AWS you can expose existing interfaces as MCP tools without changing the systems behind them. Amazon Bedrock AgentCore Gateway accepts an OpenAPI specification, an API Gateway stage, or a Lambda function with a tool schema, and presents them to agents as one MCP endpoint. It can also sit in front of MCP servers you already run, and it handles inbound authentication at the gateway (AWS).
Two design choices matter more than the tooling.
- Expose a few task-shaped tools. "Get order status" is a tool an agent can use well. Two hundred raw endpoints are a guessing game, and every extra tool is one more thing to secure.
- Write tool descriptions the way you would brief a new colleague. The model reads them to decide what to call, so a vague description produces wrong calls.
Begin with read-only tools. Add a tool that changes data only after the read-only ones have run for a while and you have looked at the logs.
When to skip MCP
MCP is one option among several. Thoughtworks has cautioned against reaching for it by default, noting that many use cases work just as well when an agent runs a local command-line tool or script (Thoughtworks). My rule of thumb: use MCP when more than one agent or client needs the same tool, or when you need central authentication and an audit trail. For one script on one machine, it adds moving parts you don't need.
Treat every MCP server as a new front door
An MCP server gives software a way to act inside your systems, so it deserves the same review as any internet-facing API. The specification is strict here. It treats the server as an OAuth 2.1 resource server that must check each token was issued for it, and it forbids passing a client's token through to a downstream service (MCP security best practices).
OWASP's MCP Top 10, still in beta, adds risks that are specific to agents. Tool poisoning hides instructions in a tool's description or output. Shadow MCP servers are ones that staff connect without anyone approving them (OWASP).
Five controls cover the basics:
- Give each tool the least access it needs, and keep read tools separate from write tools.
- Validate that every token was issued for your server, and use a separate credential for anything downstream.
- Treat tool descriptions and tool outputs as untrusted input.
- Keep an approved list of MCP servers, and block the ones that are not on it.
- Log every tool call with who asked, what ran, and what came back.
Check the data behind the tool
An agent answers from whatever the tool returns. Before you expose a system, find out who owns its data, how current that data is, and which fields an agent should never see. Stale or mislabeled data used to produce a bad report. Now it produces a confident wrong action.
A first 30 days
- Pick one workflow and the one system it depends on.
- Expose two or three read-only tools for that system.
- Put them behind a gateway with authentication and logging turned on.
- Let a small group use it, and read the tool-call logs every week.
- Add one tool that changes data, with a person approving each call.
By the end you will know what your systems need before agents can use them, which is the modernization plan you were looking for.
Exposing systems to agents on AWS?
Forti365 reviews AI workloads on AWS for spend guardrails and security at a fixed price, including the tools and MCP servers your agents can reach. Details are on the AI on AWS Guardrail and Security Review page.