Skip to content
SpecApril 3, 2026

RFC-0001: AGENTOBS — Observability Schema Standard for Agentic AI Systems

The full AGENTOBS open specification. Defines the Event Envelope, Namespace Taxonomy, Agent Span Hierarchy, HMAC Audit Chains, PII Redaction Framework, Provider Normalisation, and four Conformance Profiles (Core, Security, Privacy, Enterprise) for standardised AI observability across any LLM provider.

By S. Sriram, SpanForge

↓ Download PDF (PDF)

Overview

RFC-0001 defines AGENTOBS — an open event-schema standard for observability of agentic AI systems. It addresses the fundamental gap between general-purpose APM tooling and the requirements of production AI systems: structured cost attribution, PII redaction at source, tamper-evident audit chains, and a coherent span tree for multi-step agent runs.

What This Specification Covers

AGENTOBS standardises nine interconnected capabilities:

  1. Event Envelope — a typed, immutable container applied uniformly to every observability event, carrying identity, routing, and security fields alongside a namespace-specific payload.
  2. Namespace Taxonomy — thirty-six event types across eleven domains (trace, cost, cache, eval, guard, fence, prompt, redact, diff, template, audit), each with a typed payload schema.
  3. Agent Span HierarchySpanPayload, AgentStepPayload, and AgentRunPayload types that represent multi-step runs as an OpenTelemetry-compatible span tree, including ReasoningStep and DecisionPoint sub-types.
  4. Token and Cost ModelTokenUsage with per-category breakdowns and CostBreakdown as a typed value object — not a bare float — enabling reproducible cost attribution months after a call is made.
  5. Provider Normalisation — a language-neutral ProviderNormalizer protocol with built-in mappings for OpenAI, Anthropic, Vertex AI, Bedrock, Azure AI, Groq, Ollama, Mistral, Cohere, Together AI, and Hugging Face.
  6. HMAC Audit Chains — HMAC-SHA256 signing with prev_id linkage, verify_chain() with structured tamper detection, and key rotation without chain discontinuity.
  7. PII Redaction Framework — a five-level sensitivity scale (LOWPHI), Redactable type with annotations at construction time, and RedactionPolicy enforcement before any export.
  8. Export Abstraction — structural Exporter protocol with reference implementations for JSONL, Webhook, OTLP, Datadog, Grafana Loki, and OTel SDK bridging.
  9. Governance PrimitivesEventGovernancePolicy, ConsumerRegistry, DeprecationRegistry, and migration_roadmap() for managing schema evolution at scale.

Conformance Profiles

The specification defines four cumulative conformance profiles to enable incremental adoption:

ProfileWhat it adds
CoreEvent Envelope, ULID, spans, tokens, cost, export, OTel alignment
SecurityHMAC signing, audit chains, key rotation
PrivacyRedactable type, RedactionPolicy, pre-export redaction
EnterpriseGovernance policy, ConsumerRegistry, DeprecationRegistry, CLI compliance checks

A team can reach AGENTOBS-Core-2.0 conformance by implementing five things: Event Envelope with auto-generated ULID, SpanPayload with TokenUsage, deterministic JSON serialisation, one Exporter, and OTel attribute mapping at export time.

Specification Status

  • Status: Public Review Draft
  • Version: 2.0 (targeting v2.0 data model)
  • Schema: schemas/v1.0/schema.json (stable normative artifact)
  • Public comment period: April 3, 2026 — June 4, 2026
  • Issues: github.com/veerarag1973/Spanforge/issues
Explore more

Browse the full research library

See all resources →
The methodology

See the five-phase lifecycle in full

Explore the platform →
Talk to SpanForge

Request a briefing for your team

Get in touch →