Not “a dashboard bolted onto your agent” — it is the layer that turns agent execution into measurement and improvement. Most teams ship agents they can't tell are getting better or worse: no per-agent evals, no way to see why an answer failed, no safe path from a fixed bug to a permanent fix. AgenticPerformance instruments any agent system over OpenTelemetry, stores traces in one tenant-isolated Postgres, gates every version on a golden-set eval, clusters failures into named, trending problems, and runs a governed improvement loop — from assisted (L1) to suggested (L2) to judge-gated automatic (L3), inside a mechanically-enforced safety envelope. Engine-agnostic: it measures LangGraph, CrewAI, the OpenAI / Claude Agent SDKs, or a raw loop the same way.
Everything you need to know whether your agents are getting better — and to make them better, safely.
Instrument any agent over OTel; a normalization layer folds both OpenInference and gen_ai.* into one canonical trace model. One tenant-isolated Postgres store — no second datastore.
A mandatory deterministic baseline plus a per-agent golden set. A version gate blocks any regression against the prior version on a frozen case set; an empty golden set is a hard-fail, never a green light.
Auto-triage turns failures into stable, run-over-run clusters with durable identities and significance-gated trends — so you see real regressions, not noise.
L1 assisted → L2 suggested → L3 judge-gated automatic, inside a mechanically-enforced envelope: a diff allowlist, a content guard, and a fully-justified, rollback-able improvement ledger.
LLM judges are calibrated with stratified sampling (≥50/class) and a Wilson lower bound — not point estimates — with an independent gating judge and calibration expiry.
A per-agent read model — the score curve on the current frozen case set — exposed headless as API and MCP, ready for any console or agent.
To help you with any questions that are not listed here, we offer free, no-obligation consultations
Dashboards show tokens and latency. AgenticPerformance adds accountability and improvement: per-agent golden-set evals with a CI gate, named failure clusters, and a governed loop that turns a fixed failure into a permanent, rollback-able improvement — not just a chart.
Yes — it's engine-agnostic. It instruments LangGraph, CrewAI, the OpenAI / Claude Agent SDKs, or a raw agent loop over OpenTelemetry. Adapters also ingest AgenticOps and AgenticMind telemetry into one contract.
Apache 2.0, on GitHub (Moai-Team-LLC/AgenticPerformance). The core is open; enterprise features (SSO/RBAC, audit, fleet view, on-prem) are a separate edition.
Bun and Postgres (with pgvector, vectorscale and TimescaleDB). One datastore — traces, evals, clusters and the improvement ledger all live there, tenant-isolated by row-level security. First trace in ~5 minutes.
AgenticOps runs the fleet and AgenticMind judges its answers; AgenticPerformance measures and improves what they produce — its adapters ingest their telemetry into one trace / eval contract. All conform to the Agentic Product Standard (the Evals & observability surface).
Yes. We instrument your agents, stand up the trace store and eval gate, and wire the improvement loop as part of our agentic software development work. Reach out below.
More from our agentic practice — the standard we build to, the open-source layer that implements it, and the team that ships it.
Want to know if your agents are actually getting better — and to make them better, safely? Let’s talk.