Salesforce AI agent integration now has two real paths: build with Agentforce inside Salesforce, or connect agents you build on your own models and data to Salesforce through its APIs and MCP servers. Use Agentforce when the work lives mostly in the CRM and your admins will run it; connect your own agents when the work spans many systems or depends on data and models Salesforce does not hold. Either way, most of the effort is the same engineering: identity and field-level security, clean CRM data, sandbox testing, API budgets and an audit trail of every write.
What Salesforce now lets outside agents do
Three changes over the past year matter for CRM agent plans.
- The platform is callable by agents. On April 15, 2026, Salesforce announced Headless 360 and said, in its announcement, "Everything on Salesforce is now an API, MCP tool, or CLI command." The release lists more than 60 new MCP tools and 30+ preconfigured coding skills, and names Claude Code, Cursor, Codex and Windsurf as coding agents that can work against an org.
- Hosted MCP servers are generally available. Salesforce's public hosted MCP servers repository states they are GA and lets assistants such as Claude and ChatGPT connect to Salesforce data and logic. A hands-on walkthrough from Apex Hours lists six standard servers (SObject All, SObject Reads, SObject Mutations, SObject Deletes, Data 360 and Tableau Next), all disabled until an admin turns them on.
- Model partnerships. In October 2025 Salesforce expanded its partnerships with both labs. Anthropic describes Claude as a preferred model in Agentforce for regulated industries, with traffic kept inside the Salesforce trust boundary. Constellation Research reported that the OpenAI deal brings OpenAI models into Agentforce 360 and Agentforce 360 apps into ChatGPT.
Pricing is shifting too. VentureBeat reported that Salesforce is moving Agentforce from per-seat licensing toward consumption-based pricing, so agent volume becomes a budget line you forecast, not a fixed seat count. Confirm terms with your account team before 2027 planning.
Next question: Who builds and maintains Salesforce integrations for AI agents?
Agentforce or your own agents: how to choose
Many teams will run both: Agentforce for in-CRM tasks that admins own, and custom agents for workflows that cross the CRM, warehouse, billing and support. The table below is the decision CROs, RevOps and engineering leaders work through.
A simple rule: if more than half the steps in the workflow touch systems outside Salesforce, build the agent in your own stack and treat Salesforce as one tool. If nearly every step reads or updates CRM records, start with Agentforce and add external tools as needed. Our guide to AI agents for sales walks through which sales workflows tend to fall on each side.
Next question: Who builds custom agents on your own models and data?
The Salesforce AI agent integration routes
Salesforce's agentic integration patterns guide and its integration pattern post describe the main routes. In practice there are five.
- Your agent calls Salesforce through hosted MCP servers. The fastest way for an external agent to read and update records. Tools are self-describing, and the server runs as the signed-in user.
- Your agent calls Salesforce through the REST or Bulk APIs. Salesforce's own guidance prefers APIs when the integration touches sensitive information, needs bulk processing, or is a one-off unlikely to be reused.
- Your app talks to an Agentforce agent through the Agent API. The guide describes a session flow: authenticate, create a session, send messages, close the session, with context passed as named session variables.
- Agentforce calls your tools over MCP. Agentforce can register external MCP servers in its registry. Salesforce's post on external MCP servers says authenticated servers must support OAuth 2.0. The April pattern post marks MCP and agent-to-agent (A2A) support as beta, so check release status before you depend on them.
- MuleSoft wraps existing APIs as MCP. If you already run MuleSoft, MCP Bridge can expose those APIs to agents without rewriting them.
Most production builds combine two, for example reads through MCP and writes through a narrow, typed API you control.
Next question: How do you connect an agent to Salesforce and the rest of your systems?
Identity, permissions and field-level security
The first decision is whose identity the agent uses. Get it wrong and the agent sees too little or far too much.
- Per-user identity. The agent acts as the person who asked. Apex Hours reports that the hosted MCP servers honor profiles, permission set groups, sharing rules, field-level security and validation rules, because the OAuth token belongs to the signed-in user. Salesforce's patterns guide also notes that its MCP servers support only end-user credentials.
- Integration user. For autonomous jobs with no human in the loop, Salesforce recommends a dedicated, least-privilege integration user, the narrowest OAuth scopes possible and IP allowlisting on the external client app.
- Inherited access. Salesforce's MCP post warns that per-user OAuth inherits the user's existing permissions and does not narrow them on its own. An agent acting for a sales director can do anything that director can do, including mass edits.
Practical steps: create a permission set for each agent role, hide fields the agent should never read (compensation, health data, internal notes), keep credentials in Named Credentials rather than code or prompts, and sanitize anything that flows into tool parameters, since CRM text fields can carry prompt injection. The patterns guide is explicit that credentials must not appear in tool parameters, action configurations or Apex code.
Next question: What do senior Salesforce developers cost per month?
What to let an agent read, and what to let it write
Start read-heavy and earn write access one object at a time. A sensible default for sales and service agents:
- Read freely (within permissions): accounts, contacts, opportunities, activity history, cases, products and price books.
- Write with low risk: tasks, call notes, activity logs, draft emails saved as drafts, enrichment fields marked as agent-sourced.
- Write with human approval: opportunity stage, amount and close date, lead status and owner, case status and priority, anything that changes forecast or routing.
- Do not let the agent write: deletes, merges, ownership transfers in bulk, quote and contract terms, discount approvals, permission or sharing changes.
Apex Hours notes that delete tools sit in a separate SObject Deletes server and recommends leaving delete access off, with a Transaction Security Policy for bulk deletes. Keep that server disabled for agents.
Next question: How do you test an agent before it writes to your production CRM?
Data quality, sandboxes, API budgets and audit
Agents amplify whatever is in the CRM. Duplicate accounts, stale owners and free-text stage notes become confident wrong answers. Before go-live, work through this checklist:
- Fix the data the agent depends on. Deduplicate accounts and contacts for the target segment, make required fields required, and retire picklist values nobody uses.
- Build and test in a sandbox. Hosted MCP servers have a separate sandbox URL pattern, per Apex Hours. Seed the sandbox with realistic, masked data and run the agent against scripted scenarios, including hostile inputs in notes and emails.
- Write evals for the writes. For each write the agent can make, define what a correct update looks like and score it on every release. For Agentforce, Salesforce offers Testing Center and session tracing; for your own agents, evals live in your CI.
- Budget API calls. Every org has an API request allocation, and agents that poll or loop can burn through it. Cache reads, batch writes, use the Bulk API for large jobs and alert well before limits.
- Log every agent write. Record caller identity, session ID, the tool called, the before and after values and the model's stated reason. Salesforce's patterns guide recommends logging caller identity, session ID and request and response metadata for inbound sessions. Tag agent-made changes with a field or source so RevOps can filter them.
- Plan the rollback. Know how you will reverse a bad batch of updates before it happens.
Next question: What do AI agent developers cost per month?
A 60-day plan for one workflow
Pick one workflow with a clear owner, such as lead routing or post-call opportunity updates.
- Weeks 1 to 2: choose Agentforce or an external agent using the table above, map the objects and fields involved, set up the agent identity and permission set, and fix the data for that segment.
- Weeks 3 to 5: build in a sandbox, read-only first, then low-risk writes; write evals and the audit log.
- Weeks 6 to 8: run with a pilot team, with human approval on every high-risk write; review the log weekly and widen write access only where accuracy holds.
The team is a few senior people: a Salesforce developer who knows sharing, Flows and governor limits, an AI engineer who owns the agent and its evals, and a RevOps owner who signs off on what the agent may change.
Next question: Who can fix CRM data and automation before an agent goes live?
How Ryz can help
Ryz adds top 1% senior engineers to your team, working US business hours and reporting to your leads, and most ship in their first week. For this work that usually means senior Salesforce developers who handle permissions, data and integration, alongside an AI pod team that builds agents on your models and data inside your cloud and repos. Senior engineers typically run $10,000 to $15,000 per engineer per month, you own all code and IP, and we sign your NDA.
FAQ
Can external AI agents connect to Salesforce?
Yes. Salesforce announced Headless 360 in April 2026, exposing platform capabilities as APIs, MCP tools and CLI commands, and its hosted MCP servers are generally available. Agents can also use the REST and Bulk APIs or the Agent API.
Do Salesforce MCP servers respect field-level security?
According to a hands-on review by Apex Hours, the hosted MCP servers honor profiles, permission sets, sharing rules, field-level security and validation rules because they run as the signed-in user. Test this in a sandbox with your own permission sets before production.
Should we use Agentforce or build our own agents?
Use Agentforce when the workflow lives mostly in Salesforce and your admins will own it. Build your own when the workflow spans several systems, needs your own models or private data, or you want agent logic that is portable if the CRM changes. Many companies run both.
What should an AI agent be allowed to write in Salesforce?
Start with low-risk writes such as tasks, notes and activity logs. Require human approval for stage, amount, owner and status changes, and keep deletes, merges, pricing terms and permission changes out of the agent's reach entirely.

