Product · the six things
How Flowlines actually works.
Six surfaces, one loop. Read the traces you already store, surface what's happening across thousands of sessions, decide what to ship next.
01 · Connect
Plug into the traces you already store.
You don't send data to us. Point Flowlines at the place your traces already live, Langfuse with API keys, Helicone, Arize, or any other OpenTelemetry pipeline. Or drop JSON over a webhook.
- Langfuse · API keys · 5 min setup
- Helicone, Arize, raw OTEL · 10 min
- JSON drop over webhook · for everything else
- Backfills last 30 days on first connect
Sources · what we read from5 connected
+5mLangfuse · API keys · streamingACTIVE
+10mHelicone · OTEL endpointACTIVE
+10mArize · OTEL endpointACTIVE
+8mraw OTEL · OTLP HTTP/gRPCACTIVE
+1mJSON drop · webhook / batchREADY
02 · Explore
Walk every session, turn by turn.
Tool calls, LLM calls, memory reads, and signal annotations interleaved on a single timeline. See cost, latency, and outcome per turn, not per call. Every session has its own permalink.
- Turn-by-turn timeline with tool + LLM spans
- Inline signal pills on the offending turn
- Cost, tokens, latency per turn
- Permalink any session for Slack handoff
Sessions
12 active · 5 with signals firing · 3 clean
AllCriticalWarning
ACTIVE
12
SIGNALS
5
CLEAN
3
AVG COST
$0.04
code-agent-v3add-rate-limit-#284 · 4 turnsfabricated_test_result2m ago
support-botticket-#9482 · 7 turnsuser_frustration8m ago
tutor-englesson-12-recap · 12 turnscontext_loss14m ago
code-agent-v3refactor-#117 · 3 turnsclean22m ago
support-botticket-#9476 · 5 turnshallucination31m ago
03 · Cohort
Group users by behavior, not demographics.
Flowlines computes cohorts from observed behavior. Activity meets satisfaction: happy, neutral, at-risk. Recency meets cadence: active, dormant, churning. Or write a rule, cohorts based on prompt, tier, plan, or signal mix.
- Auto-cohorts: happy / neutral / at-risk / dormant / churning
- Custom rule-based cohorts
- Per-cohort signal mix on hover
- Power Score = CSAT × cadence
Entity overview
150 users · 4,200 sessions· last 30d
By behaviorBy plan
CSAT
Cadence
Happy but rarePower usersAt riskActive, unhappy
Satisfaction
42% happy27% neutral31% at-risk
Activity
58% active28% dormant14% churning
04 · Analyse
Before-and-after for every prompt deploy.
Every version gets a panel: signal severity, success rate, cost per session, latency. Spot a regression the same day it ships instead of next quarter's retro. Pin versions for comparison and annotate each deploy.
- Signal-by-signal before/after per deploy
- Cost, tokens, latency deltas
- Pin and compare arbitrary version pairs
- Annotations on the version timeline
Versions
Every deploy and what it measurably did
7d14d30d
Success rate %last 14d
DEPLOYS
3
MOVED THE NEEDLE
2
REGRESSIONS
1
NO EFFECT
0
05 · Patterns
Feature requests in the words your users used.
Intent extraction runs on every session. Recurring intents cluster. New intents surface as discoveries. Feature requests phrased in your users' own words show up ranked by frequency, by cohort, and by abandonment cost.
- Auto-clustered intents per session
- Discoveries queue for promotable patterns
- Cohort skew per discovered intent
- Promote any cluster into a custom signal
Signals
32 active · sorted by severity
24h7d30d
criticalwarninginfo
SESSIONS W/ CRITICAL
24%
↑ 6pt vs prior 7d
SIGNALS / SESSION
1.8
↑ 0.3
CLEAN SESSIONS
31%
↓ 9pt
All 32Critical 5Warning 8Info 19
hallucination14%47 users2m ago
cascade_failure8%22 users12m ago
user_frustration9%18 users6h ago
cost_escalation6%11 users1d ago
premature_stop3%8 users2d ago
06 · RecommendROADMAP · DESIGN PARTNER PREVIEW
Fix the root cause. Reversibly.
When a pattern is significant, Flowlines drafts a memory fix. The exact field that, if injected into the next call, would have prevented the failure across 200 sessions. Approve, reject, or auto-merge. Every write reversible.
- Typed, scoped, versioned memory fields
- Provenance back to the originating sessions
- Flag-gated rollout, one-click rollback
- Auto-merge rules for high-confidence fixes
behavior_contract.jsonDRAFTcode-agent-v3 · v14
“completion_claims”: {
−“require_tool_evidence”: false,
+“require_tool_evidence”: true,
+“evidence_tools”: [“run_tests”, “git_commit”],
}
Reviewed before merge · reversible
Ready to read your first trace?
Get early access to Flowlines. No SDK, no infra. Connect Langfuse or any OTEL endpoint and read behavior in under an hour.
Connect5 min
InstallNo SDK
Works withAny LLM