The teams running real agents on Lattice. Three of them, in their own words.
We don't ship a customer carousel that scrolls past in a banner. We pick three customers per quarter, write up what they actually do, and let their engineers say what works — and what they'd still change.
About 60% of our customers ship under NDA. The list above is the public subset. If you'd like a reference call with someone in your category, the team can usually arrange it within a week.
Replicate
Replicate uses Lattice to run their internal inference orchestrator — the system that schedules cold starts, autoscales replicas, and routes requests to the right model. Migrated four custom schedulers into one Lattice deployment in six weeks.
request ──► Lattice scheduler ──► replica pool
│ │
▼ ▼
retry / DLQ cold-start cache
│ │
└──────► trace ◄───────┘“We migrated four in-house schedulers to Lattice in six weeks. Replays alone have saved my team an entire on-call rotation per quarter.”
Granola
Granola's nightly enrichment pipeline runs for ~11 hours, processing every meeting recorded that day. Before Lattice, the team paged on retry-storm failures roughly twice a week. After: zero pages in eight months.
cron 02:00 UTC ──► Lattice run (resumable=true)
│
┌───────────────┼───────────────┐
▼ ▼ ▼
transcribe extract topics generate notes
│ │ │
└──────► write back to DB ◄───┘“Lattice is the only agent infra we evaluated that took the long-running case seriously. Our nightly runs are 11 hours. Nothing else handled it.”
Decagon
Decagon attaches Lattice evaluators to every production agent. They sample 12% of real customer conversations through a faithfulness rubric and a tone rubric, with regression alerts wired into PagerDuty. Caught two model drifts in 2025 before any customer noticed.
customer msg ──► triage agent ──► response
│
▼ (12% sample)
eval rubric · faithfulness
│
▼ (score < 0.85)
PagerDuty alert ──► oncall“The eval primitive is the thing. We catch model regressions on production traffic before customers do. That has changed how we ship.”
What our customers wish was different. Verbatim. We publish the gripes.
“The Python SDK is a year behind the TS SDK. We use Python.”
“The trace UI is brilliant for one run. Across runs, it's worse than Datadog.”
“Self-host on K8s is a chart Helm. The chart is rough.”
Want a reference call? We'll set one up.
Most of our customers will jump on a 30-minute call with a serious prospect. Tell us your category and what you're trying to ship; we'll match you with someone close to your shape.