Arize AX vs PetalTrace

Arize AX is a platform for improving AI agents by turning production behavior into investigations, evaluations, and experiments. Signal surfaces important issues, helps identify causes, and prepares fixes for review. Managed Agents can investigate failing traces, fix repository bugs, automate recurring checks, or be created from scratch. Swarm Observability tracks status and performance across an agent swarm in real time. Agent-as-a-Judge uses telemetry to produce scalable evaluations that can surface unknown failures, while end-to-end experiments help validate changes. Integrations include OpenAI, Anthropic, Google GenAI, Amazon Bedrock, LangChain, LangGraph, LlamaIndex, CrewAI, and Vercel AI SDK; tracing follows OpenInference conventions and OpenTelemetry. AX Free costs nothing and includes 10 Signal issues/month, 25k trace spans/month, 1 GB ingestion/month, and 15-day retention. AX Pro costs $50 a month, with 25 Signal issues/month, 50k trace spans/month, 10 GB ingestion/month, and 30-day retention. Enterprise pricing is custom. Self-hosted deployment keeps observability data, authentication, APIs, and UI within customer infrastructure; air-gapped deployment is available.

PetalTrace is an open-source observability platform for developers examining AI agent workflows. It captures prompts and completions, tool calls, token use, costs, and execution timelines, with a React web interface plus CLI, HTTP API, and MCP server access. Full prompt capture can include system prompts, message history, tool definitions, and model responses. Users can search prompt and completion text, compare runs for structural, content, and cost differences, and replay runs with changed models or temperatures or in mocked mode. The platform accepts standard OpenTelemetry traces, including from applications that do not use PetalFlow. PetalFlow integration adds graph topology, node inputs and outputs, and replay-capable snapshots. Trace data is documented as stored in SQLite, and the software can be built from source or run from a release binary as a local daemon. Authentication is described as future functionality and disabled by default. The public repository identifies an MIT license; no commercial pricing or usage limits are stated.