No measurement, no improvement — the Standard makes eval-driven development non-negotiable, and AgenticPerformance (APL) is the layer that implements it. It instruments any agent stack — LangGraph, CrewAI, the OpenAI or Claude Agent SDKs, or a raw agent loop — over OpenTelemetry, normalizing OpenInference and gen_ai.* conventions into one canonical model. On top of that trace store it runs per-agent golden-set evals with a version gate that blocks regressions in CI, auto-triages failures into stable, run-over-run clusters with significance-gated trends, and drives a governed improvement loop — assisted, suggested, or judge-gated automatic — inside a mechanically enforced safety envelope with a rollback-able ledger. Every agent gets a headless scorecard, exposed as API and MCP.
Everything you need to measure and improve an agent fleet — from raw spans to a governed improvement loop.
Built on the OpenTelemetry GenAI conventions with a normalization layer that maps OpenInference and gen_ai.* into one canonical model — instrument once, reason about every run.
A mandatory deterministic baseline suite plus a per-agent golden set, and a version gate that blocks any new agent version that regresses against the prior one on a frozen set.
Failed runs are auto-triaged into named, stable failure clusters that persist run over run, with trends that only fire when statistically significant — no noise-chasing.
Three autonomy levels — assisted, suggested and judge-gated automatic — inside a mechanically enforced safety envelope: a diff allowlist, a content guard and a fully justified, rollback-able improvement ledger.
Agents and content-addressed, immutable agent versions — so every trace, eval and improvement is pinned to exactly the version that produced it.
A headless read model of each agent's health — eval scores, failure clusters, trends — exposed as an API and over MCP, ready for dashboards and fleet views.
To help you with any questions that are not listed here, we offer free, no-obligation consultations
No. APL is engine-agnostic: it instruments LangGraph, CrewAI, the OpenAI and Claude Agent SDKs, or a hand-rolled agent loop over OpenTelemetry. If your stack can emit OTLP, APL can measure it.
The core is Apache 2.0, on GitHub (Moai-Team-LLC/AgenticPerformance). Enterprise features — SSO/RBAC, audit, fleet view, on-prem — are a separate edition.
It is the reference implementation of the Standard's Evals & observability surface: traces, golden-set evals with a CI gate, error analysis and a governed improvement loop, exactly as the Standard prescribes them.
They're complementary: AgenticOps runs the fleet, AgenticMind gives agents knowledge & memory, and AgenticPerformance measures every run — with AgenticSelfHealingCode repairing what breaks, closing the run → remember → measure → heal loop.
A Postgres/TimescaleDB instance as the single datastore and an OTLP endpoint — the ingest server accepts standard OpenTelemetry traces, and tenant isolation is enforced with row-level security.
More from our agentic practice — the standard we build to, the open-source layers that implement it, and the team that ships it.
Want eval-driven development wired into your agent stack? Tell us about your product and we'll show you what APL can measure from day one.