AI model
The AI that reads the request, reasons through it and decides the next step.
Software that's given a goal, works out its own steps and uses your systems to get the job done.
A person, a new email or a schedule gives it a task.
It reads the request and looks up what it needs.
It works out what to do, based on what it found.
It does the step, such as drafting a reply or updating a record.
If the task isn't finished, it goes round again.
The AI that reads the request, reasons through it and decides the next step.
Its role, your business rules and how it should behave.
The documents and records it is allowed to search and use.
Connections that let it look things up or take action in your other systems.
What it remembers during a task, and from past tasks where that's set up.
Limits on what it can do, and when a person must approve first.
Follows steps you set in advance. Something triggers it, and it runs the same sequence every time.
Decides the steps as it goes, based on what it finds. It can handle requests that are different each time.
Most businesses start with automation and add an agent only where a task needs judgement. The two often work together.
Agent loop
An agent repeats a simple cycle: decide the next step, use a tool, check the result. It stops when the task is done, a limit is reached or a person needs to step in.
Repeat or finish
The software that runs this loop is often called the agent harness. It sends the task and context to the model, runs the tools the model asks for, checks them against permissions, returns the results and keeps logs.
It also manages how much context the model sees, retries and time limits. Claude Code and OpenAI Codex each include their own harness, and Microsoft Copilot Studio offers a choice of three. This is one reason the same model can behave differently in each.
MCP
The Model Context Protocol (MCP) is one common way to connect an AI application to other software. A system provides an MCP server, which tells the AI what it can look up or do.
It is one option, not a requirement. Agents can also connect through a system's usual integration or API. With or without MCP, each connection still has to be set up, signed in and given the right permissions.
Anthropic introduced MCP in November 2024 as an open standard. Since December 2025 it has been governed by the Agentic AI Foundation, part of the Linux Foundation.
An MCP server can offer tools (actions), resources (data) and prompts to an MCP client inside the AI application. Sign-in is optional in the standard; servers reached over the internet usually use its OAuth 2.1-based authorisation. Either way, the permissions you grant decide what the agent can do.
RAG
Before answering, the system searches your documents for relevant passages and gives them to the AI. This helps answers reflect your own policies and prices.
If set up to, it can show the source of each answer. It does not guarantee the answer is correct, so important answers still need checking.
Retrieval-augmented generation (RAG) works in two stages. First, documents are split into passages and indexed, often as embeddings in a vector database. Then, when a question arrives, the most relevant passages are retrieved and added to what the model is given.
Answer quality depends on the documents themselves, how they are split, the search settings and keeping the index up to date. Access should match who is allowed to see each document.
Multi-agent
Some jobs are split across several specialist agents, with one agent coordinating the work. This is often called orchestration.
Most businesses start with one agent and add more only when a job spans different areas.
An agent can call other agents the way it calls a tool, or hand a task over completely. When the agents run in different systems, such as one run by a supplier, an open standard called Agent2Agent (A2A) handles the hand-over.
Google introduced A2A in April 2025, and it became a Linux Foundation project in June 2025. In August 2026 it joined the Agentic AI Foundation, the same body that governs MCP. Each agent publishes an Agent Card describing what it can do and how to reach it. A2A connects agents to each other; MCP connects an agent to tools and data.
Guardrails
Guardrails are the rules and checks that stop an agent doing something it shouldn't. They limit which systems it can reach, check what goes in and out, and block risky actions.
Common guardrails include giving the agent only the access it needs, filters that block personal data or off-topic requests, checks for prompt injection (hidden instructions in emails or web pages), and limits on retries and spending.
Several layers are usually combined, because no single check catches everything. Activity logs record what the agent did so issues can be traced.
Human in the loop
You decide which actions an agent can take by itself and which need a person to review first. Approvals have to be set up. An agent only asks where you've told it to.
Approval steps are usually built into the agent loop or workflow. The agent pauses, sends the proposed action to a person, for example in Teams, Slack or email, and continues only once it is approved. Each decision should be logged.
Good candidates for approval are actions that are hard to undo, such as sending to customers, making payments or deleting records.
Often, yes. An agent can connect through built-in integrations, APIs or MCP. Each connection needs the right permissions, and some systems need custom work.
It can. Depending on the setup, an agent starts when a staff member asks, when something happens such as a new enquiry arriving, or on a schedule.
Yes, where the product supports approval controls. For example, an agent can prepare a quote and wait for approval before sending it. Test which actions need approval before going live.
Pause it, review what happened and fix any affected records. To reduce the risk, limit its access, keep action logs and require review for important actions. Start with tasks where mistakes are easy to spot and fix.
Costs include setup, connections to your systems, software subscriptions or AI usage, hosting and support. The number of tasks, amount of information processed and complexity of the connections affect the total.
We help you choose and set up the right AI tools for your business.