Splyntra
Get demo
Observability · Tracing & Replay

Every autonomous agent reasoning step, traced and replayable

Traditional APMs see only single HTTP request/response cycles. Splyntra captures the full non-deterministic agent run: multi-agent handoffs, recursive tool calls, state drift, and prompt token costs on a continuous OpenTelemetry timeline.

Start free on this feature →View documentationBook technical demo
agent_trace.py
python
from splyntra import instrument, trace_agent

# Auto-instruments LangGraph, CrewAI, OpenAI Agents
instrument()

@trace_agent(name="researcher_agent")
def run_agent(task_prompt: str):
    # Every LLM invocation, tool call, & sub-agent handoff
    # streams directly into Splyntra's OTLP collector
    response = agent_graph.invoke({"input": task_prompt})
    return response
< 2ms
Client Overhead
Non-blocking OTLP streaming worker.
100%
OTel GenAI Compliant
Native semantic conventions specification.
0 Token Loss
Span Accuracy
Per-span exact prompt & completion counts.
100x Replay
State Reproducibility
Re-execute historical runs with modified inputs.

Engineered for high-throughput autonomous agents

Every capability is built into the OpenTelemetry streaming pipeline with sub-millisecond ingestion overhead.

01Graph UI

Hierarchical Span Tree & Graph Visualization

Visualize complex DAG execution graphs, LangGraph state transitions, and CrewAI task delegations on an interactive, zoomable timeline.

  • ✓Parent-child span relationships with sub-agent boundaries
  • ✓Latency waterfall breaking down tool vs LLM inference time
  • ✓Model parameters (temperature, top_p, model name) per span
02Telemetry

Exact Prompt & Completion Inspector

Inspect exact system prompts, chat history arrays, tool payloads, and JSON outputs for every reasoning loop.

  • ✓Structured view of input tokens, output tokens, and reasoning tokens
  • ✓Automatic PII and API secret masking before render
  • ✓Formatted Markdown, JSON, and raw text viewers
03Replay Engine

Interactive Agent Replay & Time Travel

Step backward and forward through multi-turn agent runs to see how context window state mutated after each tool return.

  • ✓Re-run individual sub-spans in isolation with new model prompts
  • ✓Export trace state directly into an Evaluation dataset
  • ✓Compare failed runs against successful baseline traces
Architecture & Data Flow

How it works under the hood

Native OpenTelemetry ingestion, zero-overhead client sidecars, and real-time evaluation pipelines.

Step 1OTLP Exporter

OpenTelemetry SDK Ingestion

Your agent emits standard OTLP/gRPC spans with GenAI semantic conventions.

Step 2Splyntra Core

Stream Pipeline & Token Attribution

Splyntra calculates latency percentiles and attaches per-model pricing.

Step 3Risk Engine

Risk & Security Evaluation

Inline DLP and prompt injection filters evaluate every span payload.

Step 4Dashboard

Live Dashboard & Replay

View real-time trace waterfalled trees and inspect variable states.

1-Line Instrumentation

Instrument Any Agent in 2 Lines of Code

Python (LangGraph / CrewAI)TypeScript / Node.js
from splyntra import instrument

# Automatically hooks into OpenAI, Anthropic, LangChain, LangGraph, CrewAI
instrument(
    api_key="splyntra_live_...",
    project="customer-support-agent"
)

Frequently Asked Questions

How is AI agent tracing different from standard APM tracing (Datadog/NewRelic)?▼
Traditional APMs treat requests as deterministic microservice hops. AI agents involve non-deterministic reasoning loops, dynamic tool selection, multi-agent handoffs, and LLM context window mutation. Splyntra captures token counts, prompt caching, tool input/output JSON schemas, and per-trace security risk scores natively.
Does Splyntra cause latency in my production agent?▼
No. The Splyntra SDK uses an asynchronous, background worker thread with ring-buffered batching. Span export is completely detached from the critical path of your agent's execution.
Can I export traces to standard OpenTelemetry collectors?▼
Yes. Splyntra is 100% compliant with OpenTelemetry GenAI semantic conventions. You can ingest from any standard OTel collector or forward traces outward to your existing data lake.

Related Platform Features & Guides

SecurityRisk Score & Prompt Defense →FinOpsCost Intelligence & FinOps →ObservabilityTools & RAG Retrieval →IntegrationsFramework Integrations →

Ready to monitor and secure your AI agents?

Get started in under 3 minutes with zero credit card required. Free tier includes up to 5 projects and community telemetry.

Start Free on CloudRequest Custom Enterprise Demo