Research: Init Intelligence AI ITSM Engineering Stack

Overview

The strongest engineering architecture for Init Intelligence is not “LLM chatbot plus integrations.” It is a governed operating system for IT work: intake, context graph, living playbooks, agent planning, deterministic execution, policy/approval checks, human escalation, audit evidence, and system-of-record sync.

The web research reinforces the existing wiki thesis: Init Intelligence should build toward AI-native managed outcome delivery rather than a generic workflow builder. The engineering stack should make agents useful while preventing unbounded agent behavior in sensitive IT systems.

Key Findings

• Workflow builders are converging around AI authoring, blocks, integrations, and self-hosting. Sim AI and n8n show the visual/no-code version; Windmill shows a more code-grade internal automation model; Pipedream shows reusable integration components. initlabs-engineering-sim-ai-workflow-automation-2026-04 initlabs-engineering-workflow-tooling-snapshot-2026-04
• Durable execution is mandatory for IT workflows. Onboarding, offboarding, access requests, compliance evidence, and remediation often need retries, waits, approvals, and event-driven continuation. Temporal, Trigger.dev, Hatchet, and Inngest are the main patterns to study. initlabs-engineering-durable-execution-engines-2026-04
• The workflow-vs-agent distinction should be an architecture rule. Anthropic, LangGraph, and Mastra all point to the same split: workflows for known control flow; agents for open-ended tool selection and reasoning. Production IT actions should default to workflows, not fully autonomous agent loops. initlabs-engineering-agent-frameworks-observability-2026-04
• MCP is a protocol layer, not a governance layer. MCP exposes tool schemas and calls, but Init Intelligence still needs tenant-specific permissions, credential scoping, human approvals, policy decisions, audit logs, rate limits, and execution traces. initlabs-engineering-agent-tool-governance-authz-2026-04
• Agent tool governance is a first-class product primitive. Arcade, Composio, OpenFGA, OPA, Cedar, and E2B show pieces of the tool-auth, policy, and sandboxing puzzle, but none are a complete AI ITSM governance product. initlabs-engineering-agent-tool-governance-authz-2026-04
• The context graph must ingest real IT systems, not just documents. CSDM/CMDB, SCIM, SAML, Okta, Entra, Intune, Jamf, NetBox, Backstage, Jira Service Management, ServiceNow, and compliance tools define the practical data surface. initlabs-engineering-enterprise-it-integration-substrate-2026-04
• Observability must become customer trust. OpenTelemetry and agent SDK tracing are useful, but Init Intelligence needs product-level traces: what the agent saw, what it proposed, which policy allowed it, who approved it, which tool ran, what changed, and what evidence was written. initlabs-engineering-agent-frameworks-observability-2026-04

Recommended Architecture

The product should be built as six layers:

1. Intake layer — Slack/Teams/email/portal/API; normalize requests into typed work objects.
2. Context layer — integration-fed graph of users, devices, apps, groups, roles, tickets, services, ownership, policies, and evidence.
3. Knowledge/process layer — living playbooks, runbooks, historical tickets, KBs, workflow templates, service catalog.
4. Agent planning layer — classify, ask clarifying questions, retrieve context, draft plans, propose tools/workflows, generate tests.
5. Deterministic execution layer — durable workflow runtime with typed steps, idempotency, approvals, retries, rollback notes, and system-of-record sync.
6. Governance/observability layer — policy engine, relationship authorization, tool gateway, credential scoping, audit traces, metrics, evals, evidence.

Build-vs-Buy Lean

• Build internally: request model, context graph schema, ITSM-specific approval/audit/evidence model, policy UX, workflow review surface, customer-visible trace, managed outcome operating loop.
• Consider buying/composing early: durable execution, sandbox execution, OAuth/tool auth for commodity SaaS, observability plumbing, vector/search infrastructure, low-level auth primitives.
• Do not outsource the core: the agent-tool governance layer and IT context model are likely defensible product primitives for Init Intelligence. ^[inferred]

Core Concepts

• ai-workflow-substrate — authoring, integration, execution, versioning, approval, and observability layer.
• deterministic-agent-runtime — agents plan, but typed workflows execute.
• agent-tool-governance — policy, authorization, credentials, approval, and audit for tool calls.
• integration-and-context-layer-for-ai-itsm — integration-fed context for safe AI ITSM actions.
• context-graph — live graph replacing stale CMDB assumptions.
• mcp-backed-workflow-generation — tool/schema discovery for safer workflow generation.
• ai-itsm-readiness-debt — product opportunity around cleaning operational data/process debt.

Entities & Tools

• Workflow UX and automation: sim-ai, n8n, windmill, pipedream
• Durable execution: temporal, trigger-dev, hatchet, inngest
• Agent orchestration: langgraph, mastra, openai-agents-sdk
• Governance and tool access: arcade-ai, composio, openfga, opa, cedar, e2b
• Context and observability: opentelemetry, netbox, backstage

Contradictions & Open Questions

• Visual builder vs code-grade artifact. Sim/n8n-style UX is accessible; Serval/Windmill-style code artifacts are more reviewable. Init Intelligence likely needs both: no-code review surface, code-grade execution underneath.
• Internal MCP vs public MCP. Public MCP can become a distribution channel, but internal MCP/tool-schema discovery may be enough for early workflow generation quality.
• Temporal-class runtime vs startup-speed runtime. Temporal is the reference durable execution model, but Trigger.dev/Inngest/Hatchet may move faster for the first product.
• Buy tool router vs build tool gateway. Arcade/Composio could accelerate integration/auth, but the most strategic surface for Init Intelligence may be a proprietary IT tool gateway with context, policy, and evidence built in.
• Graph visibility. The context graph is mandatory internally, but whether customers should see it directly or only through outcomes/traces remains open.

Sources Consulted

• initlabs-engineering-sim-ai-workflow-automation-2026-04
• initlabs-engineering-workflow-tooling-snapshot-2026-04
• initlabs-engineering-durable-execution-engines-2026-04
• initlabs-engineering-agent-frameworks-observability-2026-04
• initlabs-engineering-agent-tool-governance-authz-2026-04
• initlabs-engineering-enterprise-it-integration-substrate-2026-04
• initlabs-thesis-and-wedge
• agent-first-itsm-back-office-automation