Open-source Agent observability stack extracts beliefs directly from model streams
mosesman831 · reddit · 2026-08-15
The author open-sourced a stack for agent observability focused on extracting the agent's "beliefs" (assumptions, intentions, plans) directly from the streamed output, rather than just recording traces. This aims to explain why an agent failed. The stack consists of four MIT-licensed repos:
- Axion: Middleware that extracts beliefs as structured events (e.g., intention confidence 0.80) with zero added latency. Includes PII redaction and webhooks for Langfuse/Arize integration.
- VisReplay: Records full sessions (thoughts, tool calls, errors) into versioned files for frame-by-frame replay.
- VisCompile: Deterministic behavioral diffing between agent runs (byte-exact), usable in CI pipelines for regression gating.
- VisBoard: Shared agent/human workspace with versioned notes, SSE events, and scoped auth workflows.
Currently self-hosted only. Belief extraction is pattern-based (regex + clause rules) and designed to be free and deterministic.
More from coding & agent
- QualCoder MCP: Open-Source Tool for Conversational Qualitative Data Analysis with Claude — nicotem · 2026-08-15
- Anthropic reveals internal benchmark for automated AI research — sachdh · 2026-08-15
- Developer Shares Custom Diagram Skill, Installable with One Command — cneuralnetwork · 2026-08-15
- Agent-Reach: CLI tool gives AI agents free internet access across platforms — PrajwalTomar_ · 2026-08-15
- Microsoft updates Generative AI for Beginners course with RAG guide — bibryam · 2026-08-15
- Vercel v0 API goes GA: App gen enters production workflows — krishnan · 2026-08-15