Straight answers on agents in production.

You connect Flowlines by pointing it at the trace store you already use, Langfuse, Helicone, Arize, any OpenTelemetry pipeline, or a JSON drop. There's no SDK to ship and no code change; the first connect backfills your last 30 days and then streams new sessions live.

No. Flowlines reads from your existing trace store (Langfuse, Helicone, Arize, or any OTEL pipeline) instead of shipping an SDK into your agent. Setup is a five-minute read connection with zero code change.

Individual signals start firing within minutes of connecting, because the first connect backfills your last 30 days of traces. Cross-session views like cohorts and version comparisons stabilize once you have a few hundred sessions per cohort.

Any of them. Because Flowlines reads traces rather than your code, the model and framework don't matter, OpenAI, Anthropic, open-source models, LangChain, LlamaIndex, CrewAI, LangGraph, Vercel AI, or a custom loop all work the same way as long as their traces land in Langfuse or an OTEL pipeline.

An agent is working in production when its behavioral signal rates are low and stable, its cohorts skew happy rather than at-risk, and its latest version didn't regress against the previous one. Flowlines surfaces all three on top of the traces you already store, so 'is it working' becomes a number instead of a hunch.

Agent drift is the gradual divergence of an agent's behavior from its original instructions, learned user state, or expected outcomes. You detect it by comparing behavioral signal rates across rolling windows and across prompt versions, not by inspecting individual traces.

The failures that matter most are behavioral, not exceptions, the agent returns a 200 and a confident answer that's wrong, looping, or off-policy. Flowlines detects these as signals across sessions and surfaces recurring patterns you never wrote a rule for through a discoveries queue.

Flowlines ships a registry of behavioral signals that fire automatically on every session, including hallucination, cascade_failure, user_frustration, cost_escalation, low_quality_response, session_abandon, policy_violation, and high_latency, plus custom signals you promote from recurring patterns in the discoveries queue.

Most user feedback is invisible. Your users tell your agent what they want in their own language, and that signal is buried in transcripts that nobody reads. Flowlines extracts intent from every conversation, classifies it across recurring patterns, and segments users by what they actually asked for.

Frustration shows up in language before it shows up in churn: re-phrasing the same request, negative sentiment, repeated corrections, giving up mid-task. Flowlines detects this as the user_frustration signal on every session and rolls it up per user, so sustained frustration becomes a leading churn indicator instead of a postmortem.

Surveys capture the loud minority who click a thumb; the silent majority's satisfaction is inferable from behavior. Flowlines derives a CSAT proxy from in-session signals, frustration, abandonment, repeat contacts, resolution, and rolls it into cohorts (happy, neutral, at-risk) and a per-user Power Score so you can track satisfaction continuously instead of sampling it.

Flowlines groups users into cohorts computed from observed behavior rather than demographics: satisfaction buckets (happy, neutral, at-risk) on one axis and activity buckets (active, dormant, churning) on the other. You can also define custom cohorts with a rule over any traced attribute, and every signal and version view can be filtered by cohort.

Users abandon when the agent fails them in a way your logs call a success: an unanswered intent, a hallucinated answer, a loop, or a slow response. Flowlines detects abandonment as the session_abandon signal, ties each one back to the intent and failure that preceded it, and shows which cohorts abandon most.

Feature requests don't arrive as tickets, they arrive as half-finished sentences inside agent conversations. Flowlines extracts them from every transcript, clusters duplicates, and ranks by frequency and abandonment cost so you can prioritize without reading 10,000 sessions.

Evals test the cases you thought of, on inputs you curated, before you ship. Production is the cases you didn't think of, on real users, over time. Evals are necessary but they're a pre-flight check, not a flight recorder, they can't see drift, real-user intent, or the failure modes that only emerge at scale.

You answer that by comparing behavior before and after the deploy on the same real traffic, not by eyeballing a few outputs. Flowlines pins every signal rate, success rate, cost, and latency to the deployed prompt version, so each release gets a before/after panel that shows what it measurably did.

Silent failures are the sessions that return a 200 and a confident answer that's wrong, empty, looping, or off-policy, no exception, no alert. Flowlines catches them as behavioral signals (hallucination, cascade_failure, low_quality_response, and more) scored on every session, so the failures that never threw an error still surface.

You fix next whatever costs the most: the signal firing on the most sessions, the cohort closest to churning, or the unmet intent driving the most abandonment. Flowlines ranks failures by reach and cost so prioritization is a sorted list, and (in preview) drafts the memory fix that would prevent the top one.