How We Build Agents
A single model that tries to do everything rarely does anything well inside an enterprise. Real work spans your CRM, ERP, data warehouse, and a dozen bespoke systems. So we build agents the way you build teams: specialists with clear roles, governed access to systems, grounded in your data, and reviewed before anything reaches a customer.
This is the method behind every Cirrius build, whether the agent runs on Agentforce 360, Claude, or OpenAI. Salesforce itself now ships as MCP tools and APIs, and Claude runs inside Agentforce’s own trust boundary, so the agent can live in Lightning, Slack, or Claude and still act under your existing permissions. You own the code, the runbook, and the architecture memo at the end of it.
Multi-Agent Teams
Instead of one AI that reasons, searches, calls APIs, and writes the reply, we assemble a small team. Each agent has a narrow job, its own permissions, and its own logs. Agentforce agents slot in as specialists; Cirrius orchestrates around them. Each specialist’s playbook is packaged as a portable Agent Skill, an open standard that runs in Claude, Codex, or Copilot and is versioned like code.
Planner
Breaks a business goal into steps and hands each one to the right specialist.
Knowledge
Pulls facts, policies, and history from your knowledge bases and data warehouse.
Action
Executes governed operations: a work order in ITSM, an entitlement in CRM, an inventory check in ERP.
Quality
Reviews outputs for accuracy, compliance, and tone before they reach a customer or a system.
Guard
Enforces access scopes and routes high-risk actions to a human for approval.
Why a team beats a lone genius: narrow scope means fewer errors, independent tasks run in parallel, one failing agent can be retried without taking the system down, and each agent gets only the permissions its job needs. No single AI with the keys to everything. Agents reach systems over MCP and hand work to each other over A2A, both now governed by the same Linux Foundation body, so an Agentforce specialist and a Claude specialist can work the same case.
Two Examples From Live Systems
Field Service triage and scheduling
- Planner reads the fault report and sets the plan.
- Knowledge agent pulls the asset’s service history and manuals.
- Action agent checks parts and lead times in the ERP.
- Scheduler proposes technicians and appointment windows.
- Quality agent drafts the customer update; Guard routes anything over threshold to a dispatcher.
Agent Assist for the service desk
- Planner classifies the case and decides which specialists to wake.
- Knowledge agent grounds the answer in Salesforce Knowledge and past cases.
- Action agent updates the case, entitlement, and next-best-action fields.
- Quality agent checks tone and policy before the rep sees the suggestion.
- Every step is logged per agent so the team lead can see exactly what happened.
MCP Integration
Agents need to reach your systems without holding your API keys. A Model Context Protocol (MCP) server is the gatekeeper: it exposes approved data and actions to the agent through a standard, schema-driven interface, validates every request, enforces your security policy, and executes on the target system. Point-to-point AI integrations become one managed, governed layer. MCP is now a vendor-neutral standard under the Linux Foundation’s Agentic AI Foundation, supported natively by Claude, ChatGPT, Gemini, Copilot, and Agentforce, so a server built once serves every agent you run.
We run our own company this way: Claude has been connected to the Cirrius Salesforce org through MCP servers for a year, the same pattern Salesforce and Anthropic have since productized as Claudeforce, now in open beta. Agentforce agents can consume MCP servers directly through the org’s MCP registry and allowlist, so a server we build for Claude also serves your Agentforce agents. If you already run MuleSoft, existing APIs become MCP servers with no new code, and Agent Fabric gives you one registry for agents and tools across every vendor.
Enterprise APIs return far more than an agent can use, so we shape the response into context-efficient payloads and build to the current MCP spec: OAuth scopes, SSO-ready enterprise auth, long-running task tracking, PII redaction, and audit logs from day one. You own the server code.
Data and Context Engineering
An agent is only as good as what it can retrieve. Prompt engineering has become context engineering: deciding what enters the model’s window, what is compressed, what is retrieved on demand, and what is left out. Retrieval-Augmented Generation (RAG) is the retrieval half of that, grounding answers in your approved sources instead of the model’s memory. It is where most agent projects quietly fail, so we engineer it deliberately.
Prepare the knowledge
Catalog the sources, clean and chunk the documents, choose embeddings, and load a vector store alongside your structured data. Data 360 (formerly Data Cloud) is the Salesforce-native path when your knowledge already lives there, and its own MCP server now exposes that context to Claude or any MCP client.
Find the right context
Hybrid search across vectors and keywords, filtered by the user’s permissions, ranked and trimmed to fit the context window. We measure retrieval quality with evals, not vibes.
Answer with citations
The current Claude or OpenAI model, or Agentforce, writes the response from the retrieved context and cites its sources, so a reviewer can verify any answer that matters.
Typical Salesforce uses: grounding Agentforce Service replies in Knowledge and past cases, account research for sellers, SOP and policy retrieval for operations, and turning technical manuals into an instant expert for field teams. Updating the knowledge base is faster and cheaper than retraining a model, so the system stays current.
Security and Evals
Whether your security team will allow an agent to act is a separate question from whether it can. We treat it as part of the build, not someone else’s problem.
- Threat modeling and data classification for every agent and every MCP connection before code is written
- Least-privilege scopes per agent, PII redaction on the way in and out, and audit logging your CISO can read
- Human-in-the-loop policy: which actions run alone, which need approval, and how escalation works
- Evals for retrieval accuracy, task success, and tone, run before go-live and on a schedule against live traffic after it
- Every model call, tool call, and retrieval is a traced step using OpenTelemetry GenAI conventions, so debugging and tuning are practical, not guesswork
- Agentforce agents wire into Agentforce Observability and Agent Optimizer; custom agents emit the same traces, so both show up in one view
See It Applied to One of Your Workflows
Most engagements reach a Production Readiness checkpoint at Day 30 and a live agent your team owns by Day 90. Start with a complimentary workshop and leave with a written plan for one agent, the permission model it needs, and an honest read on whether it belongs on Agentforce or a custom stack.