Product · Agent
Four agents, one accountable system.
telepace isn't 'a chatbot doing surveys.' It's a harness of specialized agents, an event-sourced spine, and policy layers that keep them honest — designed to be inspected, replayed, and improved.
The team
Each agent has one job. That's why the system stays reliable.
Designer
Chats with you to turn a goal into an interview outline.
Consumes
goal · background · constraints
Emits
SpecUpdated event stream
Interviewer
Runs the live conversation. Probes. Tracks coverage. Knows when to stop.
Consumes
outline + running transcript
Emits
TurnRecorded · InterviewCompleted
Analyst
Embeds transcripts, clusters themes, drafts insight reports.
Consumes
InterviewCompleted stream
Emits
TranscriptEmbedded · ThemeClusterUpdated · InsightGenerated
Coordinator
Watches the campaign as a whole. Nudges, escalates, closes.
Consumes
every projection
Emits
NotificationSent · EscalationTriggered
The harness
Not just LLMs in a trenchcoat.
Every command routes through the Harness: a policy-gated dispatcher backed by an append-only event store. Nothing gets to an agent until the budget check, the PII check, and the escalation check say yes.
Because state is events, we can replay any campaign, A/B a policy against a corpus, or reconstruct exactly what the Interviewer heard on turn 27 of interview 84.
It's the engineering discipline your compliance team asks about — built in from day one.
Policies enforced by default
- Budget
- Per-campaign USD ceilings on LLM + voice minutes.
- PII
- Presidio + custom patterns. Redact before persistence.
- Escalation
- Distress / self-harm signals stop the run and page a human.
- Observability
- Every agent call has trace_id, cost, and eval score attached.