Zespan

Zespan is an AI agent reliability platform for engineering teams running agents in production. It automatically traces every LLM call, tool invocation, agent handoff, and planning span across an agent's full execution — built agent-first, not retrofitted from LLM monitoring. Setup is two lines of code: providers are auto-patched and agents auto-discovered, with no manual instrumentation.

Observe & Debug — Capture every agent run automatically with no sampling. Open any run into a full span waterfall where each LLM call, tool call, and memory lookup is timed and priced, with a live input/output/token inspector. Group related traces into user sessions to replay whole conversations. Zespan auto-discovers your delegation topology and renders it as a live agent graph with per-hop cost. Track success rate, latency, and cost across every tool, and compare models (Claude, GPT, Gemini) on p50/p90/p99, spend, and error rate.

Evaluate & Improve — Score every trace automatically with 12 built-in LLM-as-judge and rule-based evaluators, or your own criteria. Version, compare, roll back, and A/B-test prompts in production without touching code. Promote real production traces into regression datasets in one click. Replay agent scenarios against new prompts or models before shipping, and test side by side in the playground.

Protect & Control — One-click guardrails block PII leaks, prompt injection, jailbreaks, runaway loops, and out-of-budget calls. Cost Explorer breaks spend down by model, day, and token type with anomaly detection; Attribution assigns exact cost to every agent, tool, model, and user with no manual tagging. Forecast projects cost and latency forward with budget alerts. Alerts on latency, cost, error rate, and eval scores route to Slack, email, or PagerDuty, and incidents are correlated automatically with AI-written root cause and MTTR tracking.

ZespanPilot, a built-in AI ops assistant, reads live trace data, answers operational questions in plain English, and takes action — create alerts, toggle guardrails, roll back prompts, run playbooks — with one-tap undo.

Zespan is OpenTelemetry-native and integrates with OpenAI, Anthropic, Google Gemini, AWS Bedrock, Mistral, Groq, OpenRouter, and LiteLLM, plus LangChain, LangGraph, CrewAI, AutoGen, the OpenAI Agents SDK, Google ADK, LlamaIndex, Vercel AI SDK, Haystack, Semantic Kernel, and Pydantic AI. SDKs for Python and Node.js/TypeScript. An MCP server exposes live trace data to Claude Desktop, Cursor, and any MCP client. Free tier available, no credit card required; SSO/SAML and self-hosted deployment on higher tiers.