Design an AI-Native Architecture with Salesforce Headless 360

Think about an AI-native startup that should ship extremely customized, real-time funding suggestions whereas navigating advanced information safety and governance necessities. Because the architect, you might be liable for designing a system that connects this exterior agent to core buyer information with out sacrificing efficiency, compliance, or scalability. That is the place Salesforce Headless 360 is available in.

In episode 5 of Assume Like an Architect, we reconstruct the architectural pondering course of for this state of affairs. Fairly than focusing solely on the completed architectural design, strolling by means of this course of helps strengthen the psychological muscular tissues architects use to guage necessities, weigh tradeoffs, and decide the proper architectural method for headless entry and autonomous brokers.

Use the “What, How, and Why” Method

To design scalable options, we use a repeatable three-step technique to maneuver from uncooked enterprise necessities to justified architectural choices:

• What: Spotlight key phrases in enterprise necessities to paraphrase the essence of the issue into concise Excessive-Stage Necessities (HLRs). This helps make sure you clear up the proper drawback from the beginning.
• How: Align technical resolution choices on to your HLRs. Use stay diagramming to visualise choices and digital sticky notes to seize questions, assumptions, and choices. It’s critical to tie your assumptions to your choices, as a call might have to vary if an assumption is later confirmed incorrect.
• Why: Doc your justifications. Creating an Architectural Determination Report (ADR) ensures that stakeholders, and your future self, perceive the rationale behind a selected course.

What: From Enterprise Necessities to Excessive-Stage Necessities 

Within the “What” step, let’s have a look at the particular problem from episode 5. We begin with a uncooked enterprise requirement and use highlighting to point the important thing components, as proven beneath.

Paraphrase into HLRs 

By highlighting key phrases, we strip away the enterprise narrative to ensure we’re fixing the proper issues from the beginning. Primarily based on the AI-native agent state of affairs, we arrive at 5 clear HLRs:

1. Ingest real-time market alerts and generate uncooked funding theses.
2. Retrieve consumer mandates, holdings, and suitability constraints from Salesforce.
3. Consider every thesis in opposition to consumer context for portfolio match and compliance.
4. Floor customized suggestions to the advisors by means of Verity’s personal interface.
5. Escalate to a human and log in Salesforce when compliance thresholds should not met.

Word that the HLRs paraphrase reasonably than use the precise phrases from the enterprise requirement. That is important in validating with the enterprise stakeholders that you just understood the intention of their requirement.

To proceed with evaluating resolution choices, let’s transfer into the “How” step by following the path of questions, assumptions, and choices captured throughout our stay session.

How: Aligning options to your HLRs

Within the “How” step, we consider resolution choices in opposition to your HLRs. To maintain our pondering organized, we monitor questions, assumptions, and choices alongside our technical diagram.

Evaluating HLR 1: Ingesting market alerts and producing theses

We begin by defining the place the uncooked funding theses come from. Our first assumption is the analysis matching and era is very custom-made, proprietary mental property (IP). We resolve to make the most of Verity’s current, proprietary analysis engine because the supply for dealing with market alerts and producing these theses, assuming the internal workings should not related for our structure design.

Evaluating HLR 2 and three: Retrieving consumer mandates and evaluating portfolio match and compliance

Subsequent, let’s deal with HLR 2 and HLR 3 collectively as we have to extract the suitable information out of Salesforce and use it to guage for match. 

Retrieve consumer mandates, holdings, and suitability constraints from Salesforce

First, we decide which information to drag from Salesforce and the way we wish to try this.

1. Evaluating Agentforce agent necessity: Since retrieving the information required neither pure language processing nor reasoning, we decided that an Agentforce agent was pointless for this step and selected to make use of Headless 360, particularly a Mannequin Context Protocol (MCP) server as an alternative.
2. Limiting agent visibility: Assuming the necessity to tightly restrict the agent’s capabilities to stop broad visibility into delicate buyer holdings, we determined to make use of a named question to strictly restrict the objects and fields the agent might entry as an alternative of utilizing the usual sobject-reads server.
3. Implementing a person safety boundary: To make sure monetary advisors might solely view profiles for their very own purchasers, the system wanted to implement this safety boundary, driving the choice to hook up with the MCP server utilizing person context reasonably than a broad service account.

Consider every thesis in opposition to consumer context for portfolio match

For portfolio match, we assume the matching guidelines had been additionally proprietary Verity IP. Fairly than copying these guidelines and information into Salesforce, we resolve to push the Salesforce constraints straight into the Verity engine to do the matching.

Evaluating HLR 4 and 5: Surfacing suggestions and dealing with escalations

For the ultimate pair of necessities, we have to give attention to the person interface and the best way to deal with compliance failures. Since Verity is an AI-native firm, we ask how the person will work together with the information and make the idea they have already got a human-facing customized agent for his or her advisors. We determined this agent will act because the core orchestration layer, pulling constraints from Salesforce by way of the MCP server and fetching matched theses from the Verity analysis engine. 

If a compliance match fails, we want the agent to escalate and to log a problem. We ask the place to log the difficulty, and assume that utilizing the Salesforce Case object is one of the best technique for each escalation and reporting. This led to the choice to have the agent write a case again to Salesforce. Much like our learn entry choice, we ask if we should always use the usual sobject-mutations. We assume giving the agent broad rights to replace any object the person can entry is much too dangerous, as we don’t need the agent autonomously promoting holdings. This leads us to resolve in opposition to normal instruments and as an alternative construct a customized MCP server utilizing an invocable motion, making use of the precept of least privilege to securely log the case with out granting full write entry.

Why: Justify Your Design with Architectural Determination Data 

Within the “Why” step, you justify your course to stakeholders and your future self. An architect’s worth just isn’t solely in designing an optimum resolution, but in addition of their capability to clarify and justify the rationale. With the questions, assumptions, and choices we captured, we will apply the steerage within the Architectural Choices: A Human-Led, AI-Powered Method weblog submit on the best way to create an ADR with the assistance of LLMs.

We use the notes we took within the “How” step to:

• Tie assumptions to choices: A choice might have to vary if an assumption is later confirmed incorrect. For instance, if the Verity agent doesn’t have orchestration capabilities, our integration might have different instruments like MuleSoft Agent Cloth.
• Doc the “pondering work”: Whether or not you current a fancy diagram or a easy slide, crucial half is that the pondering work, together with questions, assumptions, and choices, is documented.

Skilled Q&A

Through the session, Leo Tran, Salesforce Distinguished Architect and Chief Architect of Platform Engineering, and Carl Brundage, Salesforce Distinguished Technical Architect, joined us to dive deeper into Headless 360 and MCP. Under are some key questions and solutions from the session.

Understanding Headless 360 and MCP

• If Salesforce is API-first, what’s the distinction since going headless? Salesforce isn’t simply accessed by means of a browser or system integrations anymore; it’s in Slack, Microsoft Groups, and ChatGPT. Embracing the MCP normal means reworking to be agent-first, sharing sources and prompts throughout numerous agentic interfaces. Not like conventional API integrations, which power AI to guess the best way to navigate uncooked information endpoints, Headless and MCP present exterior brokers with exact, labeled directions. These schemas empower brokers to execute enterprise processes intelligently, eliminating trial and error.
• What does an MCP server supply over conventional APIs? Conventional APIs are like a “dumb phonebook” of endpoints that an agent has to reverse engineer. An MCP server works particularly for brokers and acts as a semantic toolkit with labels, offering the agent with descriptions, schemas, and solutions on precisely how and when to make use of the instruments.
• Why use Headless 360 as an alternative of Agentforce? It relies on the place your clients are interfacing. Verity is an AI-native firm with its personal proprietary agent. That proprietary IP isn’t properly suited to maneuver into Salesforce. If Verity later wished to distribute that have by way of Slack or an internet site, Agentforce might function the distribution layer, however Headless permits them to deliver Salesforce into their established interface.

Safety, governance, and entry management

• Why use named queries as an alternative of giving the agent free SOQL entry by way of sobject-reads? For those who use the usual server, the agent can learn every little thing the linked person can learn. If you’d like the agent to do lower than the person (for instance, if a person can see financials, however the agent shouldn’t), you should use a customized MCP server with named queries. This scopes the agent’s entry to solely particular fields and objects with mounted standards.
• How do you limit delicate objects? First, limit person entry with permission units and sharing guidelines. If the person wants entry to the information however the agent shouldn’t have it, then management the APIs the agent can name. Utilizing customized MCP servers with named queries restricts the agent to doing lower than the person can, making certain strict governance.
• How can we stop a person from taking their login configuration and utilizing a private, open LLM to drag PII information by way of MCP? Admins can use a number of layers of safety to stop this from taking place.  First, limiting the Exterior Shopper Apps creation to approved personnel solely, successfully barring non-admins from connecting exterior brokers by way of MCP. Second, disabling the “API Enabled” permission for any customers who don’t require direct API entry. Salesforce has additionally retire legacy SOAP API logins to remove userid/password API primarily based logins. For customers who’ve approved entry, implement Defend Occasion Monitoring and Transaction Safety insurance policies to proactively detect and block unsanctioned information retrieval. The one really safe technique is to offer a sanctioned, enterprise-grade AI path that ensures zero information retention.

Implementation and readiness

• I at present use a conventional Gross sales Cloud setup with normal report pages. What ought to I count on when transferring to a Headless 360 structure, and what are the hurdles? Normal UI pages stay acceptable for wealthy information layouts. To arrange, determine gaps and experiment with Slackbot the place actions or information don’t but exist in Salesforce, because the agent gained’t have the ability to use them headlessly.
• How do you really implement entry restrictions for the MCP server itself? Implementation requires a layered method: use Exterior Shopper Apps to grant API entry to particular profiles, disable the API Enabled permission for unauthorized customers, and use Salesforce Setup to declaratively activate or deactivate particular MCP servers.

Closing wrap-up

An essential takeaway for any architect is that whereas we historically begin with the best performance, once you want extra granular safety controls than are supplied out-of-the-box, transferring from normal to customized MCP servers with named queries and invocable actions is the justified architectural selection.

A second essential takeaway is the worth of your documented choice log. A centralized, simply accessible report of why you made particular architectural decisions is an important asset for serving to stakeholders perceive your method. Moreover, it saves your future self from having to rifle by means of outdated emails months later as you attempt to reconstruct the reasoning behind a design.

• Subsequent Episode: Be a part of us for the following episode on June 25, the place we are going to apply the What/How/Why method to a brand new real-world problem.
• Vote on future episodes: Assist us form the way forward for the sequence. Inform us what matters you wish to see and the way we will present extra worth to you.