Agentic infrastructure needs agentic observability | ODSP933

Microsoft Developer argues that observability pipelines designed for humans manually standardizing logs, traces, and metrics are struggling as AI systems generate services and infrastructure faster than those pipelines can keep up.

Overview

The session explores a shift from deterministic, event-centric observability toward an approach where the observability layer can reason over telemetry and adapt as systems change. It frames this as a response to agent-driven workflows where behavior is stochastic, context-dependent, and can fail silently even when underlying infrastructure appears healthy.

Core concept: agentic infrastructure requires agentic observability

• Traditional observability assumes engineers can keep telemetry schemas and pipelines standardized as systems evolve.
• With AI generating services and infrastructure rapidly, that manual standardization model becomes a bottleneck.
• The proposed direction is an observability layer that can interpret telemetry and adapt continuously.

New analytical focus: understanding agent reasoning (not just events)

• The session highlights a need to analyze why an agent acted a certain way, not only what events occurred.
• This implies observability that can incorporate reasoning/context signals alongside standard telemetry.

Limits of deterministic observability with stochastic agent behavior

• Deterministic assumptions (stable behavior, predictable flows) break down when agents behave stochastically.
• This creates challenges for validation, guardrails, and deciding what “correct” behavior looks like in production.

Sampling and scalability challenges under high span counts

• The talk calls out tail-based sampling and scalability issues when span counts become very large.
• High-volume tracing can make it difficult to retain the right signals needed to diagnose agent behavior.

Silent failures in agent systems

• The session describes scenarios where infrastructure-level signals look successful, but the agent system still fails in ways that are not obvious from standard telemetry.
• This motivates observability approaches that can detect failures tied to reasoning, context gaps, or misinterpretation.

Human cognition constraints shape current observability

• Existing observability practices are framed as being shaped around what humans can realistically interpret.
• Agentic systems may require different abstractions and automated interpretation to keep pace.

Critique of MCP demos: partial context for external agents

• The session critiques MCP demos where external agents operate with partial context.
• This is used to motivate better validation and context-aware interpretation in observability for agent systems.

A new model: move LLM up, push analysis down

• The session proposes a model shift described as “move LLM up, push analysis down.”
• The intent is to have higher-level reasoning available while distributing analysis closer to where telemetry is produced/processed.

Data quality as a prerequisite for reliable agent reasoning

• Data quality is positioned as critical for agent reasoning and overall system reliability.
• The session raises questions around:
  • Guardrails
  • Validation
  • Autonomy

Session metadata

• Event: Microsoft Build 2026
• Session: ODSP933
• Level: Intermediate
• Topic area: Agents & apps