Field notes
What production behavior teaches us.
Practical guides to MCP distribution, analytics, instrumentation, and agent observability.
pick the question you are working on
What Melaya learned from its first 754 MCP tool calls
Melaya puts AI assistants to work across browsers, Android apps and workflows. See how Flowlines helps its team learn from real MCP usage.
Read the field note ↗How to Monitor MCP Servers in Production
Monitor MCP servers in production: tool-call success rate, errors, p95 latency per tool, alerts, multi-client usage, users and outcomes.
MCP analytics · 13 minMCP Analytics: How to See What Users Are Actually Doing With Your MCP Server
Measure MCP users, use cases, journeys, outcomes, unmet needs, feedback, and commercial signals instead of relying on tool-call volume alone.
Buyer's guide · 16 minBest MCP Observability and Monitoring Tools in 2026
Compare MCP observability and monitoring tools by layer: service health, tool calls, traces, users and sessions, alerts, OpenTelemetry setup and pricing.
MCP strategy · 10 minWhen Should You Build an MCP Server? A Practical Decision Framework
Decide whether your product needs an MCP server, an API integration, an agent skill, static context or browser automation, then test the business case.
MCP engineering · 17 minMCP OpenTelemetry Instrumentation: How to Instrument an MCP Server
Instrument an MCP server with OpenTelemetry in Python or TypeScript: trace tool calls, record latency, propagate context and validate telemetry.
MCP observability · 6 minHow to Add MCP Usage Analytics to Your OpenTelemetry, Grafana or Datadog Stack
Keep Grafana or Datadog for MCP server health and add usage analytics: who used your MCP, from which AI client, what they tried, and what needs action.
MCP observability · 10 minWhy Did My MCP Tool Call Fail? MCP Error Codes and Tool Errors Explained
What each MCP error code means, from -32602 to -32000 and -32001, how protocol errors differ from isError tool results, and who hit them.
MCP analytics · 9 minMCP Server KPIs: Activation, Retention and Calls per Completed Job
The KPIs that show whether an MCP server creates value: activation, weekly retention and calls per completed job, with formulas and pitfalls.
MCP analytics · 8 minWhy Agents Ignore Your MCP Tools: Unused, Confused and Overlapping Tools
Find MCP tools that are never called, picked by mistake or called over and over, and fix the names, descriptions and overlaps behind them.
MCP observability · 9 minMCP Went Stateless: Measuring Sessions, Users and Latency After the 2026-07-28 Spec
The 2026-07-28 MCP spec removed initialize and protocol sessions. How to identify AI clients, group sessions, count users and time retried calls.
MCP distribution · 15 minHow to Get Listed in the OpenAI Plugins Directory: MCP Submission Guide
Let Codex prepare your OpenAI plugin submission with Plugin Creator, then complete the MCP tests, listing, policies, and review.
MCP distribution · 11 minHow to Get Listed in the Claude Plugin Marketplace: Complete Submission Guide
Build, validate, submit and distribute a Claude plugin: public-repository requirements, package structure, review path and an MCP launch checklist.
MCP analytics · 8 minWhich AI Clients Call Your MCP Server, and How to Identify Users Safely
See which AI clients call your MCP server from per-request clientInfo, and attribute calls to users with verified token claims and pseudonymous IDs.
MCP engineering · 8 minMCP Server Logging: What to Log, What Never to Log
Where MCP server logs go, which fields to record for each tool call, what never to log, and how to tie logs to traces under the 2026-07-28 spec.
MCP analytics · 9 minHow to Price an MCP Server With Usage Data
Choose an MCP pricing model from real usage data: active users, calls per completed job, cost per job, heavy users and the tools that carry value.
MCP engineering · 9 minMCP Server Evals Before Launch, Analytics After: Catching Regressions
Test an MCP server before launch with Inspector, contract tests and tool-call evals, then compare versions in production to catch regressions.
MCP observability · 9 minHow to Monitor a Remote MCP Server: Health Checks, Metrics and Alerts
Monitor a remote MCP server with health checks, synthetic probes, per-tool metrics and alert rules that work across replicas and AI clients.
Guide · 5 minHow to monitor AI agents in production
A practical framework for monitoring production AI agents across execution, quality, behavior, users, and outcomes, with a path back to original sessions.
Evaluation · 4 minHow to evaluate AI agents in production
Evaluate production agents with outcome metrics, behavioral signals, cohorts and sampled human review, without treating an LLM judge as ground truth.
Engineering · 4 minHow to debug AI agents in production
Debug production agents by moving from an outcome gap to a recurring behavior, affected cohort, exact trace and verified release change.
Guide · 6 minAI Agent Failure Modes in Production: The Complete Taxonomy (2026)
A working taxonomy of production AI agent failure modes: false success, loops, drift, intent mismatch and cohort gaps, each shown in a real session.
Guide · 9 minThe 9 Best AI Agent Observability Tools in 2026
Compare nine AI agent observability tools across tracing, evaluation, production sessions, cross-session analysis, OpenTelemetry, and MCP workflows.
Architecture · 2 minWhat is behavioral observability?
Behavioral observability explains how an AI agent behaves across production sessions, users, intents, issues, tools, and outcomes.
Engineering · 2 minHow to detect AI agent drift in production
Detect agent drift as recurring behavior across sessions, users, intents, and MCP journeys, then inspect the affected production sessions.
Engineering · 3 minHow to connect Flowlines to production traces
Connect OpenTelemetry, Langfuse, or LangSmith to Flowlines, verify the first session, and analyze behavior across production traces.
Opinion · 2 minFlowlines and Mem0 solve different production problems
Mem0 provides agent memory. Flowlines provides behavioral observability across production sessions, users, issues, and MCP journeys.
Product · 2 minIntent engineering requires production evidence
Intent taxonomies become useful when they are grounded in complete production sessions, connected to outcomes, and reviewed as user needs evolve.
Architecture · 2 minLLMs are stateless. Production agent systems should preserve context.
A production agent needs explicit session, user, release and tool context so teams can reconstruct behavior and evaluate outcomes over time.
Reliability · 2 minThe silent failure problem in AI agents
Silent agent failures look technically successful while the user outcome is wrong, incomplete or abandoned. Here is how to detect them in production.
Product · 2 minWhy AI agents do not improve from production data by themselves
Production agents do not learn on their own. Improving them takes observable outcomes, reviewed failure patterns, controlled changes and release checks.