What your agentis actuallydoing.
Moda turns every production conversation into a cohort, a signal, or a fix, automatically. The dashboard you should already have.
7 of 10 share a “refund still pending”. One policy fix closes the cohort.
Every intent your users have. Including the ones they invented yesterday.
Live Taxonomy · Emergent Intents · The Unanswered
- Where is my refund?412
- Refund denied without reason287
- Wrong refund amount225
- “Cancel because of price hike”first seen 3d ago · no canonical answer yet42
- “App refuses to load with VPN on”first seen 5d ago · no canonical answer yet31
- “Where do I update my tax info?”first seen 4d ago · no canonical answer yet28
- “Refund missing on credit card”first seen 2d ago · no canonical answer yet19
Hierarchical clustering of every production conversation, plus continuous detection of intents your agent sees for the first time, surfaced before any of them have a canonical answer.
Read what your users are really doing.
Moda monitors how every conversation unfolds: escalation, hesitation, intent shifts, abandonment, the moment trust breaks. You see the turn your agent earned a customer, and the turn it lost one.
- EscalationTone shifted from polite to confrontational at turn 4
- Repeated rephrasingSame intent restated 3 times in different words
- Trust breakUser stopped following agent suggestions after a contradiction
- Intent shiftConversation pivoted from refund to account cancellation
- AbandonmentClosed window before resolution
Find the failure points that keep recurring.
Tool failures, dropped intents, contradictions, all grouped by pattern, ranked by impact, traced back to the conversations they affected. The same pipeline that finds the failure suggests the fix.
lookup_orderTIMEOUTprocess_refundINVALID_STATEsend_notificationRATE_LIMITEDverify_identityAUTH_EXPIREDBuilt to be called.
The dashboard isn't the product. It's one consumer of it. Every cluster, every behavior signal, every diagnosed failure is a structured endpoint, scriptable from your terminal, your CI, or wherever your agent already runs.
Drops into Vercel AI SDK, Claude Agent SDK, LangChain, Mastra, OpenRouter. Anywhere you ship calls.
$ moda clusters --top 3 --last 7d returns & refunds → failed 924 ↘ 12% suggested fix ready account & billing 1,847 ↗ 5% order management 1,203 ↗ 3% $ moda insight returns/failed --explain cluster: returns & refunds → failed volume_7d: 924 wow_delta: -0.12 root_cause: lookup_order timeout (3s, cold start) suggested_fix: timeout 3s → 8s + retry policy eval_cases: 23 conversations available for replay
Observation isn't the product. It's the foundation of an agent that learns.
Every production conversation, structured at ingest. No sampling.
Into a living taxonomy of intent. Categories, subcategories, leaves, reorganizing as your traffic shifts.
Behavior signals, tool failures, and intents that didn’t exist last week.
Root cause with the exact turn the agent broke trust, and the cluster of conversations it kept breaking the same way.
The patch. The eval case. The prompt amendment. The tool config drift that explains the failure pattern.
Patch ships. The loop measures whether the fix held, and whether the agent’s users felt it.
The loop runs continuously. Every pass narrows the gap between what your users want and what your agent can give them.
The agent your users meet next week shouldn't be the one they met today.
Moda is currently onboarding teams running real production traffic. Tell us what you're building and we'll send you a tenant.