Skip to content

Latest commit

 

History

History
104 lines (75 loc) · 6.81 KB

File metadata and controls

104 lines (75 loc) · 6.81 KB

LLMOps and AI Agent Landscape

Sentinel sits at the intersection of LLM tracing and regulated-enterprise infrastructure. This document clarifies where it fits, what it does not do, and where it is headed. For the strategic framing — evidence infrastructure for the regulated AI era, the Trace / Attest / Audit / Comply formula, the four modules — see docs/vision.md.


Current landscape (2026)

Category Examples What they do Sentinel
LLMOps / Tracing Langfuse, LangSmith, Phoenix, Helicone Debug LLM calls, latency/cost metrics, prompt iteration, evals Complementary — does not trace individual LLM calls or provide dev debugging; LangFuse ships a Sentinel-compatible integration
AI Agent Frameworks LangGraph, CrewAI, AutoGen, Semantic Kernel Build agent workflows and state machines Complementary — does not build agents, wraps them for policy + evidence
AI agent governance (closed beta) Dome Systems Runtime control layer for enterprise agent fleets Adjacent — Dome Systems is US-incorporated, structurally subject to US jurisdictional regimes including the CLOUD Act; Sentinel is the EU-jurisdictional alternative
AI agent governance (OSS) Microsoft Agent Governance Toolkit OWASP Agentic Top-10 coverage for LangChain / CrewAI / Dify Complementary for Microsoft-first shops — MS AGT is US-incorporated; Sentinel is the answer when EU-jurisdictional evidence is required
Sovereignty-first cybersecurity Cylake Hardware + software for sovereign agent environments at national scale Adjacent — Cylake is US-incorporated, positioned around sovereignty-themed cybersecurity messaging; Sentinel leads with Provability instead
Native cloud governance Azure AI Foundry, AWS Bedrock Guardrails, GCP Vertex AI Deeply cloud-integrated policy and observability Multi-cloud clasp — excellent per-cloud; loses value in multi-cloud / on-premise / air-gapped scenarios
Enterprise Platforms Proprietary full-stack AI platforms with built-in decision layers End-to-end AI deployment with vendor-controlled decision recording Alternative — auditor-grade evidence layer for deployments where proprietary platforms are jurisdictionally excluded
GRC platforms Sastrify, Enactia, Kovrr, ADOGRC Inventory, risk classification, and management of AI systems Partner candidates, not competitors — runtime layer (Sentinel) vs. inventory layer (GRC)
Prompt Management Langfuse Prompt Mgmt, PromptLayer, Humanloop Version prompts, A/B test, human-in-loop iteration No — does not manage prompts or do prompt engineering
Eval Frameworks DeepEval, RAGAS, UpTrain LLM output quality scoring, RAG evaluation No — does not score LLM quality or run evals
Observability OpenTelemetry, Grafana, Datadog General infrastructure monitoring, distributed tracing Complementary — OTel export shipped as sentinel-kernel[otel]; native JSON/NDJSON remains the primary format

Relationship key: Complementary = works alongside. Upstream = Sentinel wraps or builds on it. Downstream = consumes Sentinel output. Adjacent = different problem, no overlap. Designed for = Sentinel addresses this category's gap.


What Sentinel does

Sentinel occupies a category with a distinctive shape: evidence infrastructure for the regulated AI era. It produces a structured, append-only, auditor-grade artefact for every autonomous decision — capturing what was decided, under which policy, by which system, and under whose jurisdiction.

This is complementary to observability and LLMOps, not competing with them. A team can use Langfuse for prompt debugging and Sentinel for compliance records. They answer different questions:

Question Answered by
Why is my agent slow? LLMOps / observability
Is this prompt performing well? Eval frameworks
What did my agent decide, and can I prove it to a regulator? Sentinel

Integration points

Sentinel wraps agent function calls. It does not require replacing your existing stack.

If you use Sentinel integrates via
LangChain sentinel-kernel[langchain] — shipped callback handler
CrewAI / AutoGen sentinel-kernel[crewai] / [autogen] — shipped task hooks
Haystack sentinel-kernel[haystack] — shipped component wrapper
LangFuse sentinel-kernel[langfuse] — shipped sovereignty panel
OpenTelemetry sentinel-kernel[otel] — shipped span exporter
Prometheus sentinel-kernel[prometheus] — shipped textfile collector
FastAPI / Django / Jupyter [fastapi] / [django] / [jupyter] — shipped middlewares
Any LLM provider Model metadata recorded in trace (set explicitly)
OPA / Rego policies LocalRegoEvaluator — in-process, no network
Python rule functions SimpleRuleEvaluator — zero dependencies

LangGraph and PydanticAI integrations are explicitly postponed — see the Intentionally postponed section of the v3.1.0 CHANGELOG entry.


Current flow

Agent → sentinel.trace() → SQLite / NDJSON trace
                ↓ (optional)
           OTEL → Langfuse / Grafana

Traces are always written to local sovereign storage first. OpenTelemetry export is shipped as an optional extra (sentinel-kernel[otel]) and runs downstream of the local write — the local record is the source of truth and cannot be gated by the exporter.


Why a separate layer

Observability tools record what happened. Sentinel records what was decided — and attaches policy evaluation, jurisdictional metadata, and human override records that make the trace legally meaningful under EU AI Act Article 12.

The distinction matters in regulated environments:

  • An observability trace says "the model returned this output in 312ms."
  • A Sentinel trace says "this agent decided to approve this request, under this policy, which returned ALLOW, with inputs hashed as SHA-256, stored on EU-DE infrastructure, at this timestamp, with this schema version."

The second is what a regulator, auditor, or court needs.

The gap in this landscape

Manifesto Chapter VIII names the gap explicitly: no other solution simultaneously offers model-spanning consistency, EU jurisdiction for evidence, open source with exit capability, and cryptographic provability in a regulation-mapped format.

Sentinel sits precisely in this intersection. LLMOps tools watch performance, not decisions. Cloud governance is locked to one cloud. Closed-beta governance platforms are US-incorporated. Sovereignty-first cybersecurity leads with the wrong noun. GRC platforms manage AI systems, they do not sit in the runtime path of a decision.

That is the category Sentinel occupies. See docs/vision.md for the full strategic framing.