Announcements
An agent says “submitted.” What actually happened?
Explore our insurance agent walkthrough, using recorded responses and synthetic scenarios to examine claimed completion.
Read articleTHE SENTRIAL BLOG
All posts
The complete Sentrial article archive.
Comparisons
12 AI agent observability tools: a practical buyer’s guide
Compare 12 observability, evaluation, and gateway platforms by workflow, then run a repeatable pilot with your own agent traces.
Read articleComparisons
Respan, Langfuse, and Sentrial: how to choose your workflow
A practical comparison framework for gateway integration, open-source control, and investigating agent outcomes with evidence.
Read articleComparisons
LangSmith vs Braintrust: test the production-to-regression workflow
Compare LangSmith and Braintrust using a concrete workflow: investigate a production trace, review the failure, and preserve it as a regression test.
Read articleGuides
How to detect false completion in AI agents
A practical evidence model for distinguishing completed, pending, failed, and unverified agent actions without confusing missing telemetry with failure.
Read articleObservability
Connect agent traces to business outcomes with OpenTelemetry
Design trace context, application identifiers, and asynchronous outcome events so investigators can distinguish execution from task completion.
Read articleEvaluation & Testing
Production AI evaluations: sampling, review, and missed failures
Build a production evaluation workflow that measures false alarms, audits missed failures, and separates biased review queues from population estimates.
Read articlePricing
What does AI agent observability cost? Build a workload model
Estimate ingestion, retention, evaluation, review, and operating costs with an explicit workload model instead of comparing headline prices.
Read articleGuides
AI agent retry loops: distinguish recovery from repetition
Use action identity, changing evidence, and terminal outcomes to separate useful retries from unresolved loops and misleading success claims.
Read articleComparisons
Datadog vs Dynatrace for AI agents: compare the investigation workflow
Both platforms document agent evaluation. Use a failed business action to compare evidence, review, coverage and operating cost.
Read articleComparisons
Dynatrace alternatives for AI agents: define what you need to replace
Separate full-stack monitoring from agent investigation requirements, then compare alternatives using the same incident and workload.
Read articleComparisons
Reviewing Langfuse for production agents: an investigation checklist
Evaluate Langfuse’s production workflow with tests for delayed outcomes, evidence completeness, reviewer corrections and repeatable experiments.
Read articleComparisons
Braintrust alternatives: compare the work between a trace and a regression
Build a shortlist around production investigation, evaluator review and reproducible experiments rather than assuming Braintrust only supports offline evals.
Read articleGuides
AI agent monitoring for startups: a first-week implementation plan
Start with one important action, verify that its evidence reaches your monitor, then add a small reviewed failure set and an actionable alert.
Read articleComparisons
Arize Phoenix vs Sentrial: compare evidence review and application changes
Distinguish Phoenix from Arize AX, then compare the workflow from a version-sensitive failure to a reviewed regression case.
Read articleObservability
Choosing an LLM observability platform: evaluate the complete workflow
Use a repeatable trial to compare instrumentation, evidence, evaluation, review and operating cost instead of relying on a universal ranking.
Read articleComparisons
Braintrust vs Sentrial: compare production scoring and evidence review
Test the full path from an asynchronous action to a reviewed evaluation and a repeatable application comparison.
Read articleComparisons
Sentrial vs Langfuse: test the same failed business action
Compare evidence capture, evaluator review and regression workflows using a concrete asynchronous action, without assuming tracing and evaluation are separate markets.
Read articleGuides
AI agent tracing: connect execution steps to the action you need to verify
Understand what a trace records, how asynchronous actions complicate it, and what evidence an investigator needs before replaying a failure.
Read articleGuides
Production agent observability: five checks before trusting the dashboard
Verify capture completeness, action identity, evaluation coverage, reviewer evidence and a working response path before treating a green dashboard as success.
Read articleGuides
LLM monitoring: connect service health, output quality and business outcomes
Build separate views of runtime health, evaluation results and authoritative outcomes, then connect them during investigation.
Read articlePricing
Langfuse pricing: estimate billable usage before comparing costs
Use dated plan terms, measured billable units and a separate evaluator-cost estimate to forecast a Langfuse workload.
Read articleEvaluation & Testing
AI evaluations: what a passing result proves and what it leaves open
Separate application tests, production evaluation and evaluator calibration so passing scores do not hide missing evidence or unseen failures.
Read articleEvaluation & Testing
Prompt A/B testing: measure outcomes without changing the judge
Keep assignment, tool conditions and evaluation rules stable so a prompt comparison measures the agent change you intended to test.
Read articleEvaluation & Testing
Agent regression testing: preserve the incident, its controls and its evidence
Turn an observed failure into a reproducible test without losing the tool conditions or accepting a score change as proof of a fix.
Read articleComparisons
Arize Phoenix vs Braintrust: compare the path from evidence to an experiment
Both platforms document evaluation. Test how your team preserves evidence, reviews disagreements and checks a change on the same cases.
Read articleObservability
AI for observability: make the investigation assistant show its evidence
Use AI to summarize and investigate traces while separating its hypotheses from verified tool outcomes and tested repairs.
Read articleComparisons
Datadog alternatives for agents: test the investigation before switching
Compare dedicated agent tools with your existing platform using a fixed incident, coverage checks and a realistic migration cost.
Read articlePricing
Datadog Agent Observability pricing: count model calls, then apply the plan
A dated pricing example shows how LLM spans, commitment terms and retention affect an agent observability estimate.
Read articleObservability
Silent agent failures: find errors that do not raise an exception
Define observable failure conditions, preserve authoritative outcomes and audit unflagged cases without inventing a universal failure rate.
Read articleAnnouncements
Sentrial is part of Y Combinator W26
A company milestone as we build tools for understanding AI agents in production.
Read article