DAISY — Entity Scoring & Agent Orchestration
“The hard problem is not adding an agent. It is deciding which agent, skill and tool should act next.”
The hard problem is not adding an agent. It is deciding which agent, skill and tool should act next.
What Made the Problem Difficult
Enterprise Salesforce environments coordinate complex, high-stakes business operations spanning quoting approvals, customer service escalations, provisioning workflows, and contract reconciliations. Incorporating LLM intelligence into these mission-critical domains required deterministic control.
Core Platform Constraint
Single-agent LLM architectures fail in enterprise CRM contexts. They suffer from context window pollution, unpredictable tool execution, hallucinations during complex workflows, and an inability to assess whether a business entity is ready for automated intervention.
- !Prevent unconstrained or hallucinatory agent mutations on sensitive enterprise CRM records
- !Routing and agent delegation must be auditable, transparent, and grounded in deterministic entity scoring
- !Zero execution of mutating actions without strict JSON schema validation and permission enforcement
- !Isolate context windows across specialized functional domains to prevent prompt bloat and cross-task degradation
- !Architect custom orchestration on top of Salesforce capabilities rather than relying on black-box chat bots
What System Was Designed
Engineered DAISY as a custom Salesforce orchestrator agent topology. Ingress requests and CRM entity states pass through a three-stage pipeline: 1) Entity Scoring evaluates data completeness, urgency, and operational risk; 2) Dynamic Multi-Agent Routing selects the ideal specialized subordinate agent with an isolated context window; 3) Governed Tool-Calling executes validated skill schemas against Salesforce enterprise capabilities with comprehensive audit envelopes.
Pivotal Architectural Choices
Key Architectural Decisions
Score-First Triage Model vs. Immediate LLM Generation
Invoking LLMs on incomplete or invalid CRM entities wastes tokens and invites hallucinated actions.
- Prompt LLM directly on raw record payload
- Static hardcoded rule triggers without AI capabilities
Ensures agents only act on entities that meet strict data readiness criteria, minimizing expensive model calls and eliminating garbage-in, garbage-out execution.
Hierarchical Multi-Agent Topology vs. Monolithic All-in-One Prompt
A single agent trying to understand quoting rules, service SLAs, and billing policies inevitably suffers context degradation.
- Mega-prompt system instruction with 30+ tools
- Autonomous uncoordinated agent swarms
Confines context windows to specific domains, ensuring precise tool calling and complete architectural observability.
What Was Actually Built
Constructed a deterministic entity scoring engine evaluating CRM record completeness, priority tiers, and risk factors.
Engineered an intent classification router delegating qualified requests to specialized subordinate agents.
Implemented a strictly typed JSON Schema tool registry defining parameter boundaries for every Salesforce platform action.
Built an enterprise governance envelope capturing structured audit trails of scores, routing decisions, and tool outputs.
Designed context isolation mechanisms that instantiate fresh, focused prompt environments per delegated task.
Deliberate Architectural Compromises
Trade-offs & Mitigations
⚖Multi-Hop Orchestration Latency vs Operational Safety
Deterministic guardrails, audit visibility, and elimination of runaway tool execution.
Slight incremental latency incurred during orchestrator scoring and routing hops.
Pre-computed scoring vectors and optimized parallel skill evaluation.
⚖Declarative Schema Tool Gates vs Freeform Autonomous Actions
Absolute guarantee that database mutations comply with platform validation rules and governor limits.
Every capability requires formal JSON schema specification and parameter mapping.
Created standardized tool scaffolding templates and reusable skill modules.
Verified Outcomes
Successfully orchestrated multi-agent workflows across complex Salesforce business entities
Eliminated hallucinated platform mutations by enforcing strict JSON schema gates at the tool boundary
Established full audit traceability for every scoring decision, agent handoff, and tool invocation
Scoring accuracy, agent latency, and transaction volume benchmarks: [VERIFY WITH KULDEEP]
System Schematic & Data Flow
DAISY evaluates entity scores, isolates context boundaries, routes to specialized subordinate agents, and governs tool execution against Salesforce capabilities.
Text alternative for screen readers: Architecture flow: CRM Entity Ingress to DAISY Orchestrator via State & Context Ingestion; DAISY Orchestrator to Entity Scoring & Routing via Triage & Capability Matching; Entity Scoring & Routing to Specialized Agents (A/B/C) via Isolated Context Delegation; Specialized Agents to Salesforce Platform Core via Schema-Enforced Tool Execution
Architectural Conclusion
“Agent architecture becomes interesting when orchestration, skills, routing and tool execution become first-class system concerns.”
- •Enterprise AI requires software architecture, not prompt improvisation. Treating scoring, routing, and tool calling as disciplined systems ensures reliability at enterprise scale.
- •A central orchestrator governing modular specialized agents provides the ultimate balance between autonomy and corporate governance.