ROLE
Evals
Tests that tell you if the agent got worse.
84
76–92 · low confidence
Evals-first workflow with CI and human review.
Datasets and evals tightly bound to LangGraph.
STACKFOR
StackFor agent evals and observability
Provisional seed stack, low confidence, not hands-on tested. Affiliate status was not an input.
Know whether the agent got better: traces, a test set, and a place that fails the build when quality drops.
Agent fanBuilderAdvancedProvisional scores
STACK SCORE
87
Mean of the role winners, compatibility +4. Separate from each tool score.
Provisional seed scores. Method seed-v0-provisional, confidence low, not hands-on tested. The band around a number is a ±8 display label, not a measured interval. Ordering uses the point estimate.
We may earn a commission if a tracking link is connected and you buy through it. It never changes ranks. In this v1 seed, tracking URLs are empty, so “Open site” goes to the vendor. Sponsorship rules.
Recommended stack
Trace every run, build datasets from failures, and gate releases on evals.
Example compatibility note, not a measured integration test. All picks accept OpenTelemetry-style traces.
ROLE
Observability
Traces and evals.
83
75–91 · low confidence
OpenTelemetry-native tracing you can run locally (ELv2).
Managed and polished.
ROLE
Monitoring
How you hear about the outage.
81
73–89 · low confidence
Ranked tools
Sorted by StackFor Score only. A tool can appear twice when it fills two roles. Affiliate programs are not on this list.
- #1promptfoo
Evals · role rank 1
84
76–92 · low confidence
Open-source CLI tests and red-teaming in CI.
Goal fit25%86Output quality20%82Ease for vibe coders15%78Pricing and free tier15%92Integrations and API10%78Momentum10%82Community sentiment5%82No cited evidence stored. seed-v0-provisional. Not hands-on tested.
Open promptfoo - #2
83
75–91 · low confidence
Open source, self-hostable, broad integrations.
Goal fit25%90Output quality20%80Ease for vibe coders15%74Pricing and free tier15%90Integrations and API10%76Momentum10%82Community sentiment5%78No cited evidence stored. seed-v0-provisional. Not hands-on tested.
Open Langfuse - #3Braintrust
Evals · role rank 2
82
74–90 · low confidence
Evals-first workflow with CI and human review.
Goal fit25%90Output quality20%88Ease for vibe coders15%80Pricing and free tier15%62Integrations and API10%84Momentum10%86Community sentiment5%80No cited evidence stored. seed-v0-provisional. Not hands-on tested.
Open Braintrust - #4Arize Phoenix
Observability · role rank 2
81
73–89 · low confidence
OpenTelemetry-native tracing you can run locally (ELv2).
Goal fit25%84Output quality20%82Ease for vibe coders15%72Pricing and free tier15%88Integrations and API10%82Momentum10%80Community sentiment5%76No cited evidence stored. seed-v0-provisional. Not hands-on tested.
Open Arize Phoenix - #5
81
73–89 · low confidence
App errors and now AI agent monitoring.
Goal fit25%86Output quality20%88Ease for vibe coders15%74Pricing and free tier15%72Integrations and API10%90Momentum10%70Community sentiment5%82No cited evidence stored. seed-v0-provisional. Not hands-on tested.
Open Sentry - #6
77
69–85 · low confidence
Datasets and evals tightly bound to LangGraph.
Goal fit25%84Output quality20%84Ease for vibe coders15%70Pricing and free tier15%64Integrations and API10%78Momentum10%76Community sentiment5%74No cited evidence stored. seed-v0-provisional. Not hands-on tested.
Open LangSmith - #7
77
69–85 · low confidence
Managed and polished.
Goal fit25%82Output quality20%84Ease for vibe coders15%70Pricing and free tier15%64Integrations and API10%78Momentum10%76Community sentiment5%74No cited evidence stored. seed-v0-provisional. Not hands-on tested.
Open LangSmith
Draft feature checklist
Goal fit is scored against these must-haves. The list is a draft rubric, not a test log.
Evals
- Datasets
- Scorers / LLM judges
- CI gates
Observability
- Traces
- Cost and latency
- Self-host option
Monitoring
- Errors outside the LLM
Questions
Are these scores from hands-on tests?
No. Every score on this page is a provisional seed estimate with low confidence. The band around the number is a display convention, not a measured interval.
Can a paid listing change who is recommended?
No. Affiliate programs and sponsored cards are not inputs. A sponsored card, when one exists, is labeled and sits outside the ranking.
NEWSLETTER
The weekly AI stack drop
New tools worth your time, by use case. One email a week. A sponsor, when there is one, is labeled.