Trace
See every agent, handoff, tool call, retry, token, and latency event in one structured run.
See it in action →Open-source · Local-first · Framework-neutral
Ship agentic AI
with evidence.
AgenticLens traces multi-agent execution, evaluates quality and behavior, finds cost and token waste, and gates releases—without requiring a hosted backend.
102tests passing
91%code coverage
0required services
1
SupervisorDelegateresearch_agent
2
ResearchRetrieveregional support data
3
SupervisorDelegatemath_agent
4
MathTool · add18 + 24 → 42
5
ReviewerFinal responsegrounded answer
Release gatePassed✓
BUILT FOR THE AGENT STACKLangGraphOpenAIAnthropicMCPAny Python agent
THE OPERATING LENSFrom opaque runs to
From opaque runs to
release confidence.
One local toolkit connects execution evidence, quality evaluation, operational performance, and optimization decisions.
Evaluate
Score outputs and execution behavior with deterministic checks or your own LLM judge.
See it in action →Compare
Measure baseline and candidate runs to catch token, latency, cost, and reliability regressions.
See it in action →Optimize
Turn evidence into recommendations for prompts, RAG, memory, tools, models, and handoffs.
See it in action →Gate
Enforce quality, pass rate, latency, cost, and failure thresholds before production.
See it in action →Export
Keep portable JSON, CSV, Markdown, Jira-ready output, and standalone HTML reports locally.
Read the docs →INTERACTIVE PRODUCT TOUR
Trace. Evaluate. Gate.
Explore the same evidence flow used by the offline LangGraph supervisor demo included in the repository.
STRUCTURED EXECUTION
Follow the whole agent graph—not just model calls.
AgenticLens records hierarchical spans for delegation, retrieval, memory, tools, validation, retries, and final responses.
- Agent and framework identity
- Parent-child execution tree
- Tokens, cost, and latency evidence
UNIFIED EVALUATORS
Bring any judge. Keep one score contract.
Combine built-in checks with Python rules, semantic scorers, classifiers, internal services, or any LLM judge.
RELEASE POLICY
Turn quality evidence into an automated decision.
Set the minimum acceptable score. The release gate updates against the observed result.
1
SupervisorDelegate · research_agent
2
ResearchRetrieve · regional data
3
MathTool · 18 + 24 → 42
4
ReviewerFinal · grounded answer
✓
Exact matchDeterministic
1.00✓
Required toolsTrace-aware
1.00✓
Latency thresholdOperational
1.00✓
Answer qualityLLM judge
0.95Average score0.95
✓
RELEASE DECISIONReady to ship
Observed 95% · Required 90%
UNIFIED EVALUATOR FRAMEWORKYour models.
Your models.
Your rubrics.
One contract.
AgenticLens does not lock evaluation to one provider. Register a trusted evaluator, reference it in a versioned suite, and route every score through the same report and release gate.
DeterministicLLM judgeSemanticSafetyRAGCustom
judge.pysuite.yaml
from agenticlens.evaluation import (
EvaluatorRegistry, LLMJudgeEvaluator
)
registry = EvaluatorRegistry()
registry.register(
LLMJudgeEvaluator(
"quality_judge", call_your_model
)
)
report = evaluate_suite(
suite, samples, registry=registry
)●Provider-neutral · Normalized score · Gate-ready
HOW IT WORKSInstrument once.
Instrument once.
Use the evidence everywhere.
AgenticLens produces portable artifacts your team can inspect, compare, share, and enforce.
Instrument
Wrap workflows and spans with a small Python API.
Observe
Capture traces, tokens, latency, cost, and behavior.
Evaluate
Run versioned suites with built-in and custom evaluators.
Ship
Compare candidates and enforce release policies in CI.
OPEN SOURCE · MIT LICENSEYour agent data
Your agent data
stays yours.
No mandatory account. No hosted backend. No data egress required just to inspect a run. Start locally and integrate on your terms.
terminal
$ pip install agenticlens
$ agenticlens inspect trace.json
Run Summary
Status succeeded
Spans 7
Tokens 86
Latency 5.2 ms
$ agenticlens gate evaluation.json
✓ Release gate passed.