Skip to main content
Agent Sentinel helps you observe, control cost, and enforce safety policies across AI agent runs—without breaking your agent when telemetry fails.

​
What you get

​
Python SDK

  • Instrumentation: @guarded_action decorator for automatic telemetry
  • Policy enforcement: Budgets, allow/deny lists, rate limits
  • Guardrails: Built-in PII detection, content moderation (incl. Gemini semantic), loop protection, idempotency replay
  • Human-in-the-loop: Approval workflows for sensitive actions
  • Intervention tracking: See what Sentinel blocks — with Gemini-enriched explanations
  • Remote sync: Background sync to platform
  • Replay mode: Zero-cost deterministic replay for debugging
  • LLM integrations: Transparent instrumentation for OpenAI, Anthropic, Grok, Gemini
  • Framework support: Native LangChain, CrewAI, AutoGen, and LangGraph integrations
  • MCP client: Model Context Protocol for LLM-platform interaction
  • Error handling: Comprehensive error types with recovery strategies

​
Platform

  • Agent discovery: Automatic registration and metrics
  • Interventions: Track policy enforcement, with Gemini-enriched reasons and remediation hints
  • Approvals: Human-in-the-loop approval workflows
  • Kill switch: Emergency org-scoped halt for an agent / run / mission
  • Prose policies: Compile English / YAML / JSON to the policy IR via POST /policies/compile
  • Policy versioning: Immutable version history, side-by-side diff, one-click rollback
  • Evals & benchmarks: Submit benchmark results, compare runs for regressions, generate adversarial scenarios
  • Activity Ledger: Complete audit trail with advanced filtering
  • WebSocket real-time: Live event streaming
  • MCP endpoints: LLM-native API for autonomous operations
  • Stats & analytics: Cost tracking and performance metrics
  • Replay analysis: Org-wide local-replay history, determinism scoring, divergence detection
  • Compliance: EU AI Act Article 14 support

​
Web Console

  • Real-time dashboard: Live monitoring with WebSocket updates
  • Interventions: Typed guardrail panels (PII / moderation / loop / idempotency) with enriched reasons
  • Approvals inbox: Review and approve sensitive actions
  • Kill Switch (/kill-switch): Emergency controls promoted to top-level
  • Runs explorer: Execution history with replay simulation
  • Replay Hub (/replay): Org-wide ledger of local-replay submissions and savings
  • Activity Ledger: Complete audit trail with export
  • Agent management: Fleet overview with performance metrics
  • Evals (/evals): Benchmark scorecards, regression diff, scenario generator (Gemini-powered)
  • Analytics: Cost dashboards, per-action token + latency drill-down
  • MCP Explorer (/mcp): Browse the platform’s tools / resources / prompts
  • Settings: Policy builder with prose authoring, version diff/rollback, evidence DAG viewer

Quickstart

Get started in 10 minutes - instrument your agent and sync to the platform.

Python SDK

Full SDK documentation - instrumentation, policies, approvals, LLM integrations.

Platform API

Platform documentation - agents, interventions, approvals, real-time updates.

Web Console

Web UI guide - dashboard, interventions, approvals, runs, analytics.

​
How Agent Sentinel fits into your agent

At runtime your agent code calls tools / functions. Agent Sentinel wraps those calls:
  • Before execution: policy engine checks allow/deny lists, budgets, and rate limits.
  • After execution: a ledger entry is appended (inputs, outputs, cost, duration, outcome).
  • Optional: a background sync thread batches ledger entries to the platform (/api/v1/ingest/).
┌────────────────────────────┐
│ Your agent code            │
│  @guarded_action(...)      │
└──────────────┬─────────────┘
               │ (policy check + timing)
               ▼
┌────────────────────────────┐
│ Local ledger (.jsonl)      │
│  .agent-sentinel/ledger…   │
└──────────────┬─────────────┘
               │ (optional background sync)
               ▼
┌────────────────────────────┐
│ Platform                   │
│  POST /api/v1/ingest/      │
└──────────────┬─────────────┘
               ▼
┌────────────────────────────┐
│ Postgres (runs + actions)  │
└────────────────────────────┘

​
Key features

Guardrails

PII, moderation, loop protection, idempotency — first-class runtime primitives.

Interventions

Track what Agent Sentinel blocks, with Gemini-enriched reasons.

Prose policies

Compile plain English, YAML, or JSON into the policy IR.

Kill switch

Emergency org-scoped halt for any agent, run, or mission.

Evals & benchmarks

CI-gated scorecards, regression tracking, Gemini scenario generation.

Replay hub

Org-wide history of local-replay submissions and dollars saved.

​
Core concepts

Actions & runs

The core data model: runs contain actions; actions have cost and outcome.

Policies

Budgets, allow/deny lists, rate limits, and remote policy sync.

Activity Ledger

Complete audit trail with compliance metadata and export.
Agent Sentinel is fail-open for telemetry (ledger writes and sync should never crash your agent), but policy enforcement is fail-closed by design (violations raise).