Skip to content
N
Booting Engineering OS
000%
All projects
Building#4

Agent Monitoring Platform

Datadog for AI agents — trace every trajectory, score quality, attribute cost-per-task, and catch failures before customers do.

Observability and evals for AI agents — see what they do, why they fail, and what they cost.

View repoAgent observability
Status
Framework-agnostic recording SDK + engine working (trajectories, cost, failure taxonomy, alerts; 14 tests). Apache-2.0.
Difficulty
Medium-High
Time to MVP
2–3 months

Role in the platform

The reliability layer. Ships standalone and as a ContextOS module; gives the Agent Marketplace its trust signals.

Overview

A platform to monitor, debug, evaluate, and cost AI agents in production: trajectory traces (reason → tool → observe steps), a failure taxonomy, an eval harness (golden tasks + LLM-as-judge), cost-per-task, and replay.

Framework-agnostic across LangGraph, CrewAI, the OpenAI SDK, and the Claude Agent SDK via an SDK + OpenTelemetry.

The problem

Everyone is shipping agents; almost no one can see what they do, why they fail, or what they cost.

LLM-call observability exists (LangSmith, Langfuse, Helicone) but true agent-trajectory observability + evals is under-served.

What it is

A recording SDK instruments an agent; spans flow over OpenTelemetry into ingest and Postgres, then surface as a trajectory viewer, an eval harness, cost-per-task, and a failure taxonomy.

Built for the family's own agents first, then sold — standalone and as a ContextOS module.

Key features

Trajectory tracing

Full reason → llm → tool trajectories, not just isolated LLM calls.

Eval harness

Golden tasks + LLM-as-judge scoring, with regression gates in CI.

Failure taxonomy

Classify and aggregate failure modes so you can fix the common ones first.

Cost-per-task

Attribute spend to tasks, agents, and model mix — not just raw token counts.

Framework-agnostic SDK

LangGraph / CrewAI / OpenAI / Claude SDK via an SDK + OpenTelemetry; replay + alerts.

Tech stack

TypeScriptOpenTelemetryPostgreSQLRecording SDKLLM-as-judge evalsFramework adapters

Why it's different

  • Agent-trajectory observability (multi-step plans, tool-call graphs) versus LLM-call logging.
  • Evals + failure taxonomy + cost attribution in one place.
  • Framework-agnostic — not tied to one orchestration library.

Roadmap

MVP

SDK + ingest, trajectory traces, basic dashboards, cost tracking, alerts.

V1

Eval harness (golden tasks, LLM-judge), failure taxonomy, replay, teams/RBAC.

V2

Framework integrations, regression gates in CI, ContextOS module, enterprise.

V3

On-prem, SSO, advanced evals, benchmarks.