Maple

Technical Reference

The AgenticArchitecture Guide.

A working technical reference for production agentic AI inside Salesforce — and outside it. Written for the architects, RevOps leaders, and CTOs who'll have to maintain whatever they buy. No sales pitch, no buzzword soup.
Take the Readiness AssessmentBook a Strategy Session

Why We Wrote This

The AI consulting category in 2026 is full of pitch decks and short on architecture. Most “Agentforce strategy” content stops at the topology diagram. This guide goes one layer deeper: how the data foundation, reasoning model, agent runtime, experience surfaces, and governance compose in production — including the Salesforce-Anthropic native integration, TDX2026 Headless 360, and the trade-offs between Salesforce-anchored and standalone Anthropic Bedrock deployments.

01

The agent is only as honest as the data underneath it.

The Data Foundation

Salesforce Data Cloud · Snowflake · Databricks · AWS · MuleSoft

Without a unified, real-time, governance-ready data foundation, every agent above it hallucinates or stalls.

  • Harmonizes customer, product, transaction, and behavioral data into a unified golden record
  • Resolves identity across systems (anonymous web sessions → known account → product user → support contact)
  • Pre-computes calculated insights (LTV, churn probability, engagement scores) so agents can act in real time
  • Enforces governance, lineage, and compliance (HIPAA, SOC 2, GDPR, FFIEC where applicable)

Common Patterns

  • →Salesforce Data Cloud zero-copy integration into Snowflake (FinTech, incumbent FS)
  • →Databricks reverse-ETL into Data Cloud (B2B SaaS lakehouse-anchored stacks)
  • →Real-time event streaming via MuleSoft into Data Cloud for immediate agent triggers

Where most deployments break: customers underspend on the data foundation, then can't understand why the agent layer hallucinates or stalls. Budget half your engineering time here.

02

The reasoning model decides. The agent runtime acts.

The Reasoning + Agent Layer

Anthropic Claude · Salesforce Agentforce · Maven AGI · AWS Bedrock

Two deployment paths.

Path A

Native, inside Salesforce

Claude runs inside Agentforce via the Salesforce-Anthropic integration. Inference happens inside Salesforce-managed VPCs. Trust-boundary path for regulated workloads.

Path B

Standalone, on AWS Bedrock

Claude runs in your AWS account. Right answer when (i) you don't have Salesforce, (ii) the workload sits outside your Salesforce data scope, or (iii) data sovereignty demands a non-Salesforce host.

  • PQL-to-SQL routing (B2B SaaS: product usage hits Data Cloud and an agent enriches and routes)
  • L1 case deflection (FinTech, HealthTech, B2B SaaS: Agentforce or Maven)
  • Compliance-aware customer onboarding (FinTech KYC, HealthTech intake)
  • Care coordination (HealthTech, with guardrailed actions across HIPAA-eligible services)
  • Risk and fraud detection with explainable reasoning (FinTech)

Maven AGI is purpose-built for autonomous L1 customer service: the deflection-with-quality-control problem. Sometimes the right answer is Agentforce; sometimes Maven; sometimes both, layered. We architect the right call per engagement.

03

Where the agent meets the customer.

The Experience Layer

Vonage · Salesforce Service / Sales / Marketing Cloud · custom UI · Slack / WhatsApp

The TDX2026 Headless 360 announcement explicitly decoupled the experience layer from the agent runtime. Agents do their work in Agentforce; the work shows up wherever the customer is.

  • Voice + SMS agents via Vonage for compliance-aware customer outreach
  • Web chat surfaces backed by Service Cloud and Maven AGI
  • Embedded agentic experiences inside the client's own product (built on Headless 360 APIs and MCP tools)

Why this matters architecturally: the experience layer is the most visible part of the system but should be the thinnest. Most of the value lives below it. If your "agentic strategy" is a chat widget, you don't have one yet.

04

Where most agent deployments quietly fail in year one.

Agentic Governance

A monthly subscription, governed by Maple

This is the layer most SIs hand off and disappear from. Production agents drift: prompts decay, retrieval logic stales, data triggers misfire, and reasoning quality slips with model upgrades. Without active governance, your AI ROI evaporates by month nine.

  • Monthly Agent Performance Audits: what each agent did, where it failed, and where to retune
  • Prompt and reasoning-engine optimization: Claude prompts, Atlas reasoning, retrieval architectures
  • Eval framework: automated regression testing on agent behavior across releases
  • Data trigger accuracy monitoring: make sure the agent acts on the right signal at the right time
  • Cost and consumption monitoring: Salesforce's new consumption-based pricing means agent efficiency directly affects your bill
  • Quarterly business reviews: measurable Agentic ROI in cases deflected, hours saved, revenue influenced

The Salesforce partner program rewrite in March 2026 makes this category structural: outcome-based partner metrics replaced points-based scoring. Governance is now what the platform itself rewards.

Reference Architectures

Same four layers,different stacks per industry.

B2B SaaS

SF + Databricks + Anthropic + Maven

See architecture

FinTech

SF FSC + Snowflake + AWS + Anthropic in trust boundary

See architecture

HealthTech

Health Cloud + AWS + Maven + Vonage + Anthropic

See architecture

What We Ask in Week One

The questions we walkthrough during discovery.

Showing this on the public page is a feature: it qualifies inbound and demonstrates seriousness.

  1. 01What's your current ratio of data to action? How much data do you have that nothing acts on automatically?
  2. 02What's the headcount-to-revenue ratio of your customer-facing teams, and how is it trending?
  3. 03Where does your team spend time today that could be automated with measurable accuracy?
  4. 04How is your data classified, governed, and secured: before we put an agent on top?
  5. 05What's your acceptable hallucination rate per use case? (And how would you know?)
  6. 06What does "agent transferred ownership to your team" look like in 18 months?

FAQ

Real questions,honest answers.

Is Anthropic the only LLM option inside Salesforce?

Today, Anthropic is the only model fully embedded inside the Salesforce trust boundary via the recent native integration. Other models can be wired in via API, but with weaker compliance posture.

What does it cost to get an agent into production?

Engagements typically range $250K–$1.5M for greenfield builds, with managed-services retainers for ongoing Agentic Governance.

Can you deploy agents without Salesforce?

Yes. We deploy Claude standalone via AWS Bedrock or direct API for clients who don't have Salesforce in their stack.

Does this work without Data Cloud?

Probably not for production. Data Cloud (or an equivalent unified data foundation) is the foundation agents need to be honest. We've seen many "agent" projects fail because the data underneath isn't there.

Ready to architectyour agentic stack?

Take the Readiness AssessmentBook a Strategy Session