COMMON PLATFORM
One AI platform. Multiple specialized agents.
The six products share common services so client integrations can be added without rebuilding the agent core.
- Sales Qualification
- Customer Support
- Document Processing
- Operations Workflow
- CRM Intelligence
- Legacy Modernization
Reference architecture
How a request travels through the stack.
A call enters at the versioned API edge, is authenticated and scoped to a tenant, reaches the orchestrator, acts through the tool layer and lands in the data and observability layer as an auditable record.
Client UI / External System
layer 1Your product, portal, workflow tool or backend service
API Gateway / BFF
layer 2Versioned /v1 surface for every agent
Authentication + Tenant Context + Validation
layer 3OAuth 2.0 or scoped API keys, server-derived tenant_id, schema validation
Agent Orchestrator
layer 4Where the agent decides what to do next
- Model Provider Adapter
- Prompt / Instruction Registry
- Retrieval / Knowledge Service
- Workflow / State Machine
- Policy & Guardrail Engine
Tool & Integration Layer
layer 5Vendor-specific connectors, isolated from agent logic
- Tool Registry — schema + allowlist
- CRM
- Ticketing
- Email / Calendar
- Documents / Storage
- Generic REST APIs
- Webhooks
Data Layer + Observability
layer 6PostgreSQL · Vector Store · Object Storage · Queue
- Logs
- Metrics
- Traces
- Audit
- Evaluation
- Cost Telemetry
Platform features
What every agent inherits.
These are core services, not per-project add-ons. An agent gets them by being on the platform.
Multi-agent orchestration
One orchestrator runs every agent, each with its own instructions, tools and state model.
Model-provider abstraction
Providers sit behind an adapter, so a model change is configuration rather than a rewrite.
Prompt / version management
Instructions are registered and versioned, so behaviour changes are reviewable and revertible.
RAG / knowledge retrieval
Tenant-scoped retrieval grounds answers in approved sources and returns references with them.
Workflow state management
Long-running work is durable. A step that already reached an external system is never run twice, so retrying after an outage is safe.
Tool registry and permissions
Every tool declares its input schema, its impact and the scope it needs. A tool outside the workflow's allowlist is refused even though it exists.
Human approval gates
Configurable gates hold any side-effecting step until a person approves it.
Tenant isolation and RBAC
Data, configuration and credentials are separated per tenant with role-based access.
Audit logging
Inputs, decisions, tool calls and outcomes are recorded with correlation IDs.
Webhook / event support
Signed, versioned, retry-safe events keep external systems in step with agent activity.
API-first integration
Anything the UI can do is available over the API, because the UI uses the same contract.
Evaluation / regression testing
Golden cases run against prompt and model changes before they reach a client environment.
Configurable data retention
Retention windows are set per tenant and per data class rather than assumed.
Client integration readiness
Designed so your systems can be added later without a rebuild.
The reusable core stays clean; anything vendor- or client-specific lives at the edges where it can change independently.
- Vendor-specific connectors remain isolated from agent logic.
- External IDs are stored with internal IDs.
- Side-effecting actions support idempotency.
- OAuth / API keys can be supported through adapters.
- Mock connectors allow demos without client credentials.
- Client-specific configuration lives outside the reusable core.
Next step
Put the platform behind your workflow.
Bring a workflow and the systems it touches. We'll map it onto the platform and show you where the approval gates belong.
Or email sales@crewtac.com