Source: Agent Frameworks and Observability for AI ITSM
Sources: Anthropic: Building effective agents, LangGraph docs, Mastra agents, OpenAI Agents SDK guardrails, OpenTelemetry semantic conventions
What It Covers
This source distills current agent-framework guidance into engineering implications for Init Intelligence’s AI ITSM product.
Key Claims
- Anthropic separates workflows from agents: workflows follow predefined code paths; agents dynamically direct their own tool usage. Anthropic advises starting simple and adding agentic complexity only when it demonstrably improves outcomes.
- LangGraph provides stateful, long-running agent orchestration with durable execution, human-in-loop, memory, streaming, and debugging.
- Mastra is a TypeScript agent/workflow framework: agents handle open-ended tool use; workflows define structured step control flow with schemas, branching, suspension/resumption, and streaming.
- OpenAI Agents SDK includes guardrails for input, output, and tool calls, plus tracing for LLM generations, tool calls, handoffs, guardrails, and custom spans.
- OpenTelemetry semantic conventions provide a common language for traces, logs, metrics, events, resources, and spans across polyglot systems.
Implications for Init Intelligence
- The product should default to workflows for production IT actions and use agents for discovery, triage, drafting, routing, and bounded tool selection. Fully autonomous agents should be reserved for low-risk or heavily constrained actions. ^[inferred]
- A production AI ITSM runtime should emit traces for:
- intake,
- model calls,
- context retrieval,
- policy decisions,
- tool proposals,
- approvals,
- external API calls,
- durable workflow steps,
- human overrides,
- final ticket/evidence updates.
- The observability model should be customer-visible enough to support trust and auditor review, not just developer debugging.
Limitations
- Agent frameworks move quickly; current framework APIs may change faster than ITSM buyer requirements.
- OpenAI and LangGraph tracing defaults are not a substitute for product-owned audit trails and evidence records.