A Real-World Agentforce Lens on the Well-Architected Framework Pillars

How can AI improve productiveness throughout our 10,000-person world skilled companies group? Our World Supply Structure staff has spent the final 18 months answering that query. We started by embedding generative AI into our circulate of labor and later grew to become one of many first Salesforce groups to construct a collaborative, multi-org Agentforce expertise internally. All through this journey, from preliminary idea to a stay system, the ideas of the Salesforce Properly-Architected Framework offered our north star, guaranteeing that our options had been trusted, straightforward, and adaptable. Now, we wish to share what we’ve discovered, utilizing the framework to construction our key takeaways.

Undertaking overview

Agentforce, a part of the Salesforce Platform, offers highly effective instruments to construct autonomous brokers rapidly, permitting architects to give attention to the important thing to enterprise success: guaranteeing that brokers are trusted, straightforward to make use of, and adaptable. This requires a stable architectural method to deal with distinctive concerns, corresponding to non-deterministic conduct, new knowledge entry patterns, and pure language interfaces.

On this article, we’ll study the Properly-Architected Framework’s three pillars by way of the lens of the Agentforce implementation that we did in-house for the Salesforce Skilled Providers group. Our implementation will function a real-world instance, consisting of three specialised brokers, a Information Administration Agent, a Scoping Agent, and a Undertaking Administration Agent. They’re deployed throughout two Salesforce orgs uncovered by way of a number of channels, together with our inner Salesforce orgs, an Expertise Cloud website, and Slack, and supported by Information 360 and MuleSoft. See the diagram beneath.

Structure diagram of the in-house Agentforce implementation for the Skilled Providers group.

Properly-Architected Pillar 1: Trusted

The Trusted pillar of Properly-Architected calls for that options be safe, compliant, and dependable. Architects make this precept a actuality by way of design decisions in areas like knowledge classification and failover, which immediately defend buyer knowledge and assure enterprise dependability. When constructing our Agentforce implementation for Skilled Providers, belief was our prime precedence. We subsequently engineered the system particularly to satisfy the safety, compliance, and reliability calls for of an enterprise-scale, autonomous Agentforce platform.

Safe

A safe agent is constructed on a three-tiered “defense-in-depth” technique: first, by classifying knowledge for automated safety; second, by choosing a safe processing boundary; and third, by guaranteeing that the agent’s permissions are match for function.

• Information classification: Step one is the implementation of field-level knowledge classification within the Salesforce metadata. Setting the Compliance Categorization on delicate fields offers specific directions to the Einstein Belief Layer, the safe gateway for all LLM interactions on the platform. This classification is the first management for Computerized Information Masking, the place the Belief Layer intercepts and redacts delicate knowledge earlier than it’s despatched to an LLM. Our Information Administration Agent wanted to have the ability to reply questions on staff and their expertise, and this required us to establish which fields we thought of PII knowledge, for instance Title and E mail handle and to replace our metadata fields knowledge classification accordingly. 
• Belief boundary patterns: With knowledge classification in place, the number of an applicable belief boundary is a crucial architectural selection. Every sample represents a definite design:
  • Shared belief boundary: Greatest for lower-risk use circumstances the place entry to exterior fashions is a precedence. The safety of this sample depends completely on knowledge classification and the Belief Layer’s masking capabilities. That is the sample we chosen for the Information Administration, Scoping, and Undertaking Administration brokers, ensuring that we had our knowledge classifications in place. Moreover, we additionally made positive that any delicate knowledge contained in any unstructured knowledge we despatched to the LLM was cleaned from the information themselves previous to sending them throughout. For instance, we leveraged a MuleSoft structure to wash any information contributed with delicate info in them.
  • Salesforce-hosted boundary: The beneficial, secure-by-design choice for brokers that deal with delicate knowledge or have strict knowledge residency necessities, as all knowledge and processing stay throughout the Salesforce ecosystem.
  • Carry Your Personal Mannequin (BYOM): Offers most management for highly-regulated industries, however shifts the duty for securing and managing the mannequin’s infrastructure to the implementing group. That is the sample that we’ll be including for predictive use circumstances that we’re exposing by way of our brokers, permitting us to coach our personal mannequin for predictions based mostly on our personal knowledge.
• Agent context: Relying on whether or not the agent is made accessible inside an authenticated context or in an unauthenticated context, totally different brokers sorts with their very own knowledge entry concerns apply, for instance:
  • Agentforce Worker Agent (AEA): A key architectural benefit of the unified Salesforce Platform is that an worker agent inherits the person’s permissions because it runs within the person’s context, routinely respecting all object permissions, field-level safety, and sharing guidelines. Provided that we’re in an authenticated Salesforce org situation for our Scoping and Undertaking Administration brokers, that is the agent sort we picked for each of those brokers. 
  • Agentforce Service Agent (ASA): These brokers run within the context of their very own designated person (which is similar for all finish customers of the agent). Which means the actions that this agent has entry to should be sure that they scope the information entry appropriately. For our Information Administration Agent, we initially picked this agent as a result of we expose our information base on an Expertise Cloud web site as initially this was the one mannequin supported in a MIAW arrange. We’re shifting to an AEA nevertheless, in order that we will extra simply expose our agent in Slack.

Compliant

Past safe, a trusted agent should additionally function in a compliant method, adhering to authorized and regulatory boundaries. This requires verifiable controls to show adherence to rules like GDPR or HIPAA.

• Information residency: The agent’s belief boundary sample is the first management for assembly knowledge residency necessities. For probably the most stringent wants, architecting the answer on the Salesforce-hosted boundary inside a geo-specific Hyperforce occasion is the verifiable selection. For our Agentforce occasion, a key architectural choice was to construct in our present Salesforce orgs, which reside within the U.S. Our choice was based mostly on a threat evaluation of our preliminary use circumstances, which we labeled as decrease threat. For future use circumstances involving stricter knowledge residency rules, our technique may lead us to utilizing the Salesforce-hosted boundary inside a geo-specific Hyperforce occasion.
• Observability: The Einstein Belief Layer is central to compliance enforcement and offers two features which might be crucial for auditors:
  1. It enforces a contractual Zero-Retention coverage with exterior LLM suppliers, guaranteeing that shopper knowledge is just not saved or used for coaching.
  2. It creates an immutable audit path of each immediate, response, and knowledge interplay, which is then saved in Information 360. This log offers the required proof to reply any auditor’s questions on who accessed what knowledge and when.

After we first launched our Information Administration Agent, there have been restrictions for Information 360 1 that didn’t permit us to make use of the Information 360-based observability knowledge. We initially solved that by logging all questions and solutions in our Salesforce org to have the required traceability. Now, nevertheless, observability is absolutely supported for companion orgs, and we’re switching to the information that’s logged in Information 360 and uncovered by way of Agentforce Studio.

Moral

By way of moral requirements, customers ought to by no means be misled into pondering they’re interacting with a human. They need to even be made conscious that content material is being generated and may not be correct. After we deliberate to permit our Information Administration Agent to assist discover consultants inside our Skilled Providers group, we requested permission from the HR staff to have the ability to question and present worker expertise previous to constructing and releasing this performance. 

Dependable

For an agent to be dependable, the agent’s structure should account for the non-deterministic conduct of LLMs and for inevitable technical failures. On prime of that, one of many key predictors for perceived reliability is the accuracy and worth of the information that the agent has entry to with a view to produce its output.

When designing for sleek failure of our Agentforce implementation, we thought of the next elements:

• Automated restoration: The usage of clever LLM failover handles supplier outages, offering a primary line of protection that reroutes requests to make sure enterprise continuity. We aren’t at the moment leveraging this performance for our Agentforce occasion based mostly on our threat evaluation of the present use circumstances. 
• Sleek degradation: When a generative job fails, the agent may fall again on a deterministic talent. As an example, if our Information Administration Agent is unable to offer a summarized reply or interpret the query, it might probably resort to a keyword-based search and return a set of outcomes. 
• Clever escalation: For intents past an agent’s functionality, the design should embrace a seamless handoff to a human that transfers the total dialog context. We initially experimented with a handoff to a specialised agent for sure varieties of queries, nevertheless, we couldn’t but move the total dialog context. In our use circumstances, we should not have entry to human brokers handy off to, as a substitute we recommend related Slack channels to proceed the dialog.
• Clear failure: As a final resort, the agent ought to fail truthfully with a transparent error message that acknowledges the issue and suggests another path.

Information high quality

After all, technical resilience is just one piece of the reliability equation. An agent is finally a mirrored image of the information it makes use of, and no quantity of failover logic can compensate for flawed info. Recognizing this, we made knowledge high quality a foundational step in our agent design. For our Information Administration Agent, this meant beginning with probably the most highly-rated information in our information base, excluding knowledge that we weren’t positive of. For our Scoping Agent, we leveraged AI-preprocessing to organize our knowledge for predictive duties, and for the Undertaking Administration Agent, we invested in constructing new knowledge factors that allow it to evaluate undertaking well being. Every agent’s efficiency is immediately tied to this knowledge funding, proving that strong knowledge governance processes are a foundational step in constructing dependable brokers.  

Properly-Architected Pillar 2: Simple

The Simple pillar ensures that options are intentional, automated, and interesting, in order that they ship tangible worth to the enterprise and finish customers. For our brokers, this meant aligning every one to a selected enterprise context to resolve the precise drawback in the precise method. This focus prevents wasted effort, improves adoption amongst our colleagues, and interprets on to the standard of the conversational expertise.

Intentional

An intentional design defines the strategic function of the agent earlier than improvement begins. It’s about making deliberate architectural decisions to make sure that the agent solves the precise drawback successfully.

• Persona-driven agent: We formally outlined personas for every of our brokers to information their tone and interplay fashion. For instance, the persona for our Information Administration Agent initially began as a “Salesforce Architect” to reflect our early customers. Because the agent’s expertise expanded to deal with extra advanced Skilled Providers questions, the persona advanced right into a “Supply Skilled”: educated, environment friendly, and collaborative. Such personas dictate how brokers work together, guaranteeing that they really feel like a useful staff member slightly than only a instrument.
• Dialog design: Leveraging our HCC staff’s dialog design method, we deliberately designed welcome messages for our brokers and explored making a extra personalised and contextualized expertise based mostly on person roles and web page location.
• Design-led method: We adopted a design-led method utilizing Parts.cloud. Earlier than writing a single immediate, we mapped the end-to-end enterprise course of that the agent wanted to assist. For instance, for our Scoping Agent, we mapped the whole course of from preliminary request to draft assertion of labor. This course of map is then used to generate the agent’s core directions, guaranteeing that the conversational circulate is immediately tied to a confirmed enterprise workflow. All three of our brokers are constructed on this basis.

Automated

An “straightforward” answer makes use of the precise automation for the precise job. The architect’s function is to put every agent functionality on the right level on the determinism spectrum — from absolutely predictable to intelligently versatile.

• Designing on the determinism spectrum: Our main architectural sample is to make use of non-deterministic prompts for understanding and interpretation, which then set off deterministic actions.
  1. Non-deterministic instance: A person asking the Undertaking Administration Agent, “What’s the standing of undertaking xyz?” requires the agent to interpret the obscure request, perceive the person’s context, and synthesize info. That is dealt with by a versatile immediate template.
  2. Deterministic instance: A request like “Add 8 hours to my timesheet for undertaking xyz in the present day” calls for a extremely deterministic motion. The agent confirms the main points after which reliably calls a normal circulate to create the time entry.
• Guardrails: Non-deterministic automation additionally implies utilizing guardrails. For deterministic flows, the logic itself incorporates the guardrails and bounds. Nevertheless, for the agent’s non-deterministic skills, the guardrails should be written immediately into its core directions. An instance of 1 guardrail we utilized is: “Solely reply questions that you could reply based mostly on context knowledge offered, don’t reply when you can not reply based mostly on the outcomes.”
• Unifying automation for end-to-end workflows: The true energy of our brokers comes from orchestrating various kinds of automation by way of the Atlas Reasoning Engine. Our Information Administration Agent exemplifies this sample. When a brand new technical doc is submitted, the agent orchestrates a multi-step course of:
  1. An AI-powered immediate summarizes the doc content material for a brand new information article.
  2. A deterministic circulate populates the abstract into the right fields of the Information Article object.
  3. A second AI-powered immediate analyzes the abstract and suggests related matter tags.
  4. A ultimate deterministic circulate applies the confirmed tags to the article file.

This offers the very best of each worlds: an clever, context-aware interplay backed by predictable and dependable enterprise processes.

Participating

An attractive agent is the direct results of an intentional, architected design. It guides customers, operates inside secure boundaries, and builds belief by way of readability.

• Create clear directions: To forestall person frustration, our brokers instantly state their capabilities. For instance, the opening message for our Undertaking Administration Agent is: “Hey, I’m your Undertaking Administration Assistant. I can present undertaking standing updates, flag dangers, or enable you log your time. What would you love to do?”
• Present guardrails: An agent should function inside an outlined scope with a view to be reliable and, by extension, participating. For our brokers, guardrails are a crucial architectural element for security and dependability. For instance, our Scoping Agent is explicitly prohibited from making monetary commitments to stop enterprise threat. If requested to vow a worth, it safely deflects the request by stating its limitations. Equally, to make sure that it’s a secure and equitable instrument, our Information Administration Agent has guardrails that stop it from answering questions on staff past their expertise profile.
• Give specific affirmation: Earlier than executing an motion that modifies knowledge, our brokers are designed to get specific affirmation. The Undertaking Administration Agent, for example, will ask: “To substantiate, you wish to add a two-week threat to the undertaking timeline for ‘Shopper UAT’. Is that right?” The agent solely proceeds after receiving an affirmative response.

Properly-Architected Pillar 3: Adaptable

The Adaptable pillar is about resilience to alter and composability, permitting options to gracefully evolve with enterprise wants. Each architectural choice includes trade-offs. Designing for adaptability means these trade-offs are made consciously, in order that brokers proceed to ship worth as necessities and applied sciences change (which they inevitably will).

Resilient

Within the context of adaptability, resilience is about efficient agent lifecycle administration (ALM). Our brokers require a course of for managing change that’s secure and scalable. We discovered a design-first lifecycle administration course of to be the simplest method. Listed below are the important thing levels on this course of:

• Design & Doc: The agent’s logic is first outlined in our course of mapping instrument, which serves because the official blueprint. When we have to add a brand new functionality, corresponding to enabling the Scoping Agent to entry a brand new knowledge supply, the method circulate is first up to date within the diagram.
• Govern & Approve: Stakeholders overview and approve the proposed adjustments towards the brand new design. This creates an audit path earlier than any configuration is touched, and as soon as authorised, we will generate new directions immediately from the diagram.
• Implement & Take a look at: With an authorised blueprint, we configure the agent in a sandbox. We use Agentforce’s Testing Middle to validate efficiency, and as a ultimate high quality gate, we deploy a beta model of the agent into manufacturing. This enables a choose group of customers to check the brand new performance towards stay manufacturing knowledge, guaranteeing that it performs as anticipated earlier than a full rollout.
• Deploy: The ultimate configuration is promoted by way of our customary, ruled CI/CD pipeline.

A design-first lifecycle wants knowledge to information it. A suggestions loop offers this knowledge, turning deployment right into a cycle of steady enchancment.

• Suggestions loop: We have now applied a suggestions mechanism immediately inside Slack, permitting finish customers to provide a easy thumbs-up or thumbs-down ranking after an interplay with the Undertaking Administration Agent. This offers rapid, actionable knowledge on response high quality. Whereas this functionality is just not but accessible in the usual Messaging for In-App and Net (MIAW) UI, we plan to undertake it as quickly as it’s launched.
• Formal governance: This suggestions is operationalized by way of a proper governance course of. Enterprise and technical stakeholders meet usually to overview efficiency analytics and suggestions patterns. This perform permits us to make data-driven choices, prioritize a backlog of recent agent expertise and corrections, and feed these priorities immediately into our ALM cycle.

Composable

A composable structure, constructed from unbiased and interchangeable elements, is central to our adaptable agentic ecosystem. Our complete structure is decoupled, leveraging MuleSoft with Information 360 and Agentforce.

• Reusable actions and logic: The inspiration of our composable design is a library of discrete actions. Nevertheless, we’ve taken this a step additional by designing a whole agent as a reusable element. Our Information Administration Agent, which lives in its personal devoted org, is engineered to be a modular, invocable supply of fact.
• Interoperability and multi-agent orchestration: Our composable structure permits our brokers to interoperate. Our Scoping Agent and Undertaking Administration Agent, which stay in separate Salesforce orgs, can name the Information Administration Agent as an motion. When the Scoping Agent wants the newest finest practices for a selected know-how, it doesn’t want its personal information base; it merely calls the Information Administration Agent to profit from its specialised expertise and knowledge. This sample permits us to construct specialised brokers that do one factor nicely and share their capabilities throughout the enterprise.
• Decoupled design: This agent-to-agent communication is feasible due to a decoupled design that separates logic from the presentation channel. This “construct as soon as, deploy in all places” technique is vital to our structure. The core logic of our Information Administration Agent will be surfaced in an Expertise Cloud website for self-service, uncovered in Slack for finish customers, or known as by one other agent, all with out refactoring its elementary design.

Conclusion

Constructing efficient autonomous brokers is enabled by architectural planning. The distinctive challenges they current — round knowledge safety, reliability, and flexibility — can’t be an afterthought. Making use of the ideas of the Properly-Architected Framework permits architects to implement the required rigor to design and construct Agentforce options which might be trusted, efficient, and designed to evolve.