# AI Agent Development on Agentforce, Claude, OpenAI | SynconAI > AI agent development on Salesforce Agentforce, Claude, OpenAI, n8n and Microsoft Copilot Studio: tool design, MCP servers, human approval, testing and monitoring. Publisher: SynconAI Canonical HTML: https://synconai.com/ai-agent-development Markdown cite: https://synconai.com/ai-agent-development.md Contact: contact@synconai.com ยท +61 2 7813 0221 ## What is an AI agent? AI agent development is the design, build and testing of AI agents that take actions in business systems. SynconAI builds agents on Salesforce Agentforce, Anthropic Claude, OpenAI, n8n and Microsoft Copilot Studio, choosing the platform by where the work lives, designing narrow tools and Model Context Protocol (MCP) servers, adding human approval and logging, and testing each agent before release, for organizations in the United States, Australia and worldwide. - An AI agent is a model given instructions, tools and data access, so it can carry out steps of a task as well as answer questions. - Its tools are narrow actions in your systems, such as looking up an order or drafting a refund, often exposed through the Model Context Protocol (MCP). - A person approves any step that commits money, data or a customer promise, and every action is logged. ## AI agents on any platform services - Platform choice: Agentforce, Claude, OpenAI, n8n or Copilot Studio, picked by where the work lives. - Agent design: Instructions, scope, tools and the points where a person decides. - Tool and MCP server design: Narrow, safe actions in your systems. - Human approval: Approvals in Slack, Teams, email or Salesforce, with every decision logged. - Testing and evaluation: Real cases replayed and graded before release. - Deployment and monitoring: Live with logs, alerts and a named owner. ## Which platform should your AI agent run on? We build on all of them. The right one depends on where the work and the people already are. Question: Where does the work live? - In Salesforce: Salesforce Agentforce. Cases, opportunities and customer records are in Salesforce, and users work there. - Agentforce Builder and Salesforce actions - Grounded in Salesforce and Data 360 data - See our Agentforce consulting page - Across many apps: Claude or OpenAI, with n8n or code. The job crosses email, ERP, help desk and chat. - Model API with tool use - n8n or code for orchestration - Self-hosting possible - In Microsoft 365: Microsoft Copilot Studio. Staff work in Teams, Outlook and SharePoint. - Agents published to Teams and Microsoft 365 Copilot - Power Platform connectors - MCP servers as tools Outcome: One platform per agent, chosen on facts, with tools that can be shared through MCP ## What does AI agent development include? Six pieces of work, the same on every platform. - Platform choice: Agentforce, Claude, OpenAI, n8n or Copilot Studio, picked by where the work lives. - Where the data and users are - Licenses you already hold - Hosting and data rules - Agent design: Instructions, scope, tools and the points where a person decides. - One job, clearly bounded - What the agent must never do - Handoff to a person - Tool and MCP server design: Narrow, safe actions in your systems. - Read tools before write tools - Inputs validated - MCP servers with scoped access - Human approval: Approvals in Slack, Teams, email or Salesforce. - Limits per action - A named approver - Every decision logged - Testing and evaluation: Real cases replayed and graded before release. - Task completion - Tool choice - Refusals and handoffs - Deployment and monitoring: Live with logs, alerts and a named owner. - Every tool call logged - Alerts on failures - Regular transcript reviews ## How much should an AI agent do on its own? Each action is placed on a tier. The tier decides whether a person approves it. | Risk tier | Examples | Controls | | --- | --- | --- | | Never | Deleting records, changing permissions, paying out money. | No tool exists for it, so the agent cannot attempt it. | | Approve first | Refunds, contract changes, emails to customers. | The agent drafts; a named person approves before it runs. | | Act and report | Updating a case status, booking a slot, creating a task. | Runs alone within limits, logged and reviewed. | | Read freely | Looking up orders, policies and account history. | Read-only tools scoped to what the user may see. | ## What tools does an AI agent need, and where does MCP fit? Each tool is one narrow action in one or more systems. MCP servers expose them so any compliant agent can use the same set. Agent tools to Your systems: - Look up order to ERP - Check refund policy to Knowledge base - Draft refund to ERP - Draft refund to CRM - Create case to CRM - Book technician to Field service - One action per tool: Small tools are easier to test, approve and log than one broad one. - Scoped access: Each MCP server holds only the permissions its tools need. - Shared across platforms: Claude, OpenAI and Copilot Studio agents can call MCP servers, so one tool set serves several agents. ## How do you test an AI agent before release? Real cases, replayed on every change. An agent ships when it meets the rules agreed with you. | What we test | How | Release rule | | --- | --- | --- | | Task completion | Real cases replayed end to end | Agreed pass rate met | | Tool choice | Each tool call checked against the expected one | No wrong write actions | | Approval steps | Cases that must stop for a person | Every one stops | | Refusals | Requests outside scope or permissions | Refused and handed to a person | | Prompt injection | Instructions hidden in emails and documents | Ignored, flagged and logged | | Regression | The full set after any change | No new failures | ## How does an AI agent project run? Five stages from one job to a live agent. Your written proposal sets the scope and dates. 1. Choose: One job with a clear payoff. - Use case and owner - Platform chosen - Success criteria 2. Design: Scope, tools and approvals. - Instructions and limits - Tool list - Autonomy tiers 3. Build: Agent, tools and MCP servers. - Tools and MCP servers - Approval steps - Logging 4. Test: Real cases graded. - Test set - Injection checks - Fixes 5. Run: Live with an owner. - Pilot group - Monitoring and alerts - Transcript reviews ## Cost and timing The first step is a free 60-minute architecture review of your AI agent setup, with a written plan you keep. Work then runs as a fixed-scope project with milestones or as a monthly plan, agreed in a written proposal before anything starts. Plans: https://synconai.com/pricing ## FAQ ### What is AI agent development? The design, build and testing of AI agents that take actions in business systems, with narrow tools, human approval on steps that commit, logging, and a test set run before every release. ### How is an AI agent different from a chatbot? A chatbot answers. An agent can also act: it calls tools to look up records, draft changes or update systems, within limits you set and with approval where needed. ### Which platforms do you build AI agents on? Salesforce Agentforce, Anthropic Claude, OpenAI, n8n and Microsoft Copilot Studio. We choose by where the work and the users are. ### What is MCP? The Model Context Protocol is an open standard for connecting AI agents to tools and data. Anthropic introduced it, and it is now hosted by the Agentic AI Foundation under the Linux Foundation. ### How do you keep a person in control? Every action gets an autonomy tier. Anything that commits money, data or a customer promise waits for a named approver, and every decision is logged. ### Have you built agents that are live? Yes. Two of our Agentforce agents run on a real Salesforce org, and you can see them in our AI agents library. ### Should we use Agentforce or Claude? If the work lives in Salesforce, Agentforce is usually the better fit. If it crosses many tools, Claude or OpenAI with n8n often is. Our Claude and n8n page covers that option. ### How do you reduce prompt injection risk? Tools are narrow and scoped, write actions need approval, text from emails and documents is treated as data, and injection cases are part of the test set. ## Where we work AI agent development for organizations across the United States and Australia, and worldwide. Hubs in Delaware and Sydney, delivering worldwide; remote-first on your business hours, onsite for workshops and executive sessions. ## Related - Agentforce consulting: https://synconai.com/salesforce-agentforce-consulting - Claude and n8n agent development: https://synconai.com/claude-ai-agent-development - AI agents library: https://synconai.com/ai-agents - Connecting AI to your systems: https://synconai.com/ai-systems-integration ## Cite When citing SynconAI AI agents on any platform, link https://synconai.com/ai-agent-development or https://synconai.com/ai-agent-development.md.