HS

Himanshu Sharma

Lead Platform Engineer

← All blogs

AI Agents • iPaaS • Knowledge Graphs

Agent Tools, iPaaS Adapters, and Knowledge Graphs

The fastest way to make agents dangerous is to give the model raw product APIs. The fastest way to make them useful is to wrap those APIs as governed iPaaS tools and ground every call in a knowledge graph of entities and policies.

Why tools should be iPaaS adapters

Freshworks already has messaging, CRM, and orchestration adapters. Agents should not grow a second integration estate. Each tool is a versioned adapter: auth, retries, schemas, and SLOs already exist.

  • One IAM story for humans, automations, and agents.
  • One CDC trail when an agent mutates a ticket or account.
  • One place to turn a tool off without redeploying prompts.

Retrieve → decide → act → write-back

1. Retrieve KG subgraph + ACL 2. Decide Planner worker 3. Act iPaaS tool task 4. Write-back CDC → graph edges Policy engine: tenant, residency, PII tags, tool allow-list blocks step 3 if the retrieved subgraph says the action is out of policy
Retrieve → act pipeline Policy gate before tools
The graph is both the read model for planning and the write model after tools run.

Tool catalog and risk classes

Every adapter is registered with a risk class that drives approval, rate limits, and whether a human must confirm before execution.

Read-only search_tickets get_account list_threads Mutating create_ticket assign_agent send_sms Restricted delete_account export_pii billing_refund Planner only sees tools allowed for tenant + residency + role
Risk class is metadata on the adapter, not a prompt hack.

What the knowledge graph actually stores

Knowledge graph work on the platform team is about modeling relationships between entities and business concepts so agents (and humans) get contextual discovery across products.

Node / edgeExampleAgent use
Account —owns→ TicketAcme → INC-4421Scope retrieval
Ticket —mentioned_in→ ThreadWhatsApp laneChannel-aware replies
AgentRun —used_tool→ Adaptercreate_ticketEvals / replay
Policy —constrains→ ToolEU residencyBlock illegal calls
Entity —similar_to→ Entityduplicate contactsDedupe before write
Graph writes go through CDC the same way CRM writes do. If the agent’s tool succeeds but the edge never appears, the saga is not complete.

Tool schema, not free-form JSON

Publish a strict JSON schema per tool. The planner may only emit that shape. Workers validate before hitting the product API. Version adapters independently of prompt packs so a Teams bot change does not require a model redeploy.

Catalog

Tool name, domain, risk class, schema.

Bind

IAM role + tenant + residency.

Execute

Conductor task, idempotency key.

Index

CDC upserts graph edges.