Supafone Labs ยท Documentation

Why Supafone#

Supafone Labs exists because production voice agents fail at system boundaries, not because developers need another prompt wrapper.

Problems we repeatedly encountered#

ProblemFailure in productionInnovation in the package
Self-supervisionThe speaking model misses cross-turn intent, emotion, workflow drift, and unsupported claimsVoice Watcher and SecondMind maintain a separate belief and directive loop off the audio hot path
Provider fragmentationEvery platform emits different transcript, tool, lifecycle, and control eventsCanonical event algebra and 14 audited adapters create one runtime contract
Weak intervention controlsA generic prompt cannot safely alter an active callCapability-aware compilation chooses native control, developer-owned context, host hook, observation, or no action
Tool hallucinationThe agent says a booking, transfer, or delivery happened before the tool confirms itTruth state separates verified outcomes from conversational language
Repeated agent engineeringEvery customer brief becomes another prompt, router, stage graph, and webhook projectAgent Factory creates an inspectable plan plus managed delivery primitives
Manual testingA few employee test calls miss the scenarios real callers produceObjective-driven adversarial QA generates cases from the actual agent contract
No stable quality scoreGeneric LLM scores fluctuate without an operational targetSSR grading converts nominal verdicts into inspectable score distributions
Scattered operationsCalls, recordings, transcripts, and decisions live in separate vendor consolesDurable activity APIs and telemetry expose one history to SDKs and product UIs
Multilingual discontinuityLanguage changes duplicate transcripts, lose state, or use the wrong voiceTranscript-authority rules and opt-in language/voice profiles preserve one canonical call
Rebuilt customer infrastructurePhone, WebRTC, SMS, campaigns, numbers, signing, and writebacks are implemented repeatedlyUnified SDK, REST, and MCP surfaces expose the surrounding operating system

The design response#

mermaid
flowchart TD
    problem[Production call problem] --> event[Canonical event]
    event --> state[Deterministic call state]
    state --> watcher[Watcher belief and directive]
    watcher --> gate[Confidence, truth, and policy gate]
    gate --> compile[Provider-aware compiler]
    compile --> action[Native action or safe no-op]
    state --> evidence[Replay, QA, telemetry, and optimization]

The architecture follows four rules:

  1. The live call remains primary. Supervision never blocks the speaking
  2. agent.

  3. Facts and warmth remain separate. Empathy cannot override tool truth,
  4. consent, or policy.

  5. Capability is explicit. Supafone never pretends a provider supports a
  6. control it does not expose.

  7. Evidence drives improvement. Calls can be replayed, graded, compared,
  8. and used to improve standing directives.

What developers stop rebuilding#

Continue with Framework coverage to see exactly how the runtime maps onto each supported platform, then use the SDK quickstart to create or supervise an agent.

View raw Markdown