Inside Current

An LLM agent with a real-time instrument.

Current does not generate audio and it does not control playback directly. It plans against a curated sound catalog; Soundriver validates those plans and translates accepted decisions into smooth, audible changes.

LLM integrationAgentic workflowStructured outputReal-time inferencePersistent stateObservability

The decision loop

From intent to audible change.

Each cycle combines probabilistic planning with deterministic enforcement. The model supplies creative judgment; application code owns safety, correctness, and execution.

  1. 01
    Listen

    Gather live context

    Current receives the active mix, session history, creative direction, listener messages, preferences, and hard constraints.

  2. 02
    Ground

    Supply the sound catalog

    The planner sees only eligible sounds and curated metadata—instrument, mood, energy, role, pitch content, and playback limits.

  3. 03
    Plan

    Ask the LLM for a decision

    The model proposes a small set of additions or removals, a public rationale, persistent direction changes, and when to decide again.

  4. 04
    Validate

    Parse structured output

    A strict schema rejects malformed commands, invented identifiers, invalid timing, and unsupported operations before they reach playback.

  5. 05
    Constrain

    Enforce deterministic rules

    Server-side policy checks source and layer limits, catalog eligibility, listener rules, harmonic compatibility, and transition safety.

  6. 06
    Act

    Execute the transition

    The browser loads audio, fades sounds together, animates the interface, and reports completed removals back to the session.

  7. 07
    Observe

    Record what happened

    Decisions, activity, messages, token usage, and state changes remain traceable for the listener and for diagnostics.

Engineering boundaries

Creative where it helps. Deterministic where it matters.

The model owns

  • Creative interpretation
  • Source-selection proposals
  • Mix direction and rationale
  • Conversation classification
  • Decision cadence proposals

The application owns

  • Catalog access and source IDs
  • Schema and policy validation
  • Musical compatibility rules
  • Session concurrency and limits
  • Audio loading and transitions

Production concerns

More than a prompt.

Context engineering

Compact session history, active rules, listener intent, and catalog metadata are assembled for each decision.

Prompt safety

Catalog text and listener input are treated as untrusted data, while the planner is restricted to supported commands.

Failure containment

Invalid or conflicting proposals stop at validation; transient inference failures do not end playback.

State and concurrency

Versioned session updates prevent stale decisions from silently overwriting newer mix state.

Evaluation

Deterministic scheduling and policy behavior are exercised through simulations and automated transition tests.

Observability

Planner decisions and usage data are recorded alongside visitor-facing activity and playback diagnostics.

Hear the system work

The architecture ends in an experience.

Enter a dynamic soundscape, watch Current explain its decisions, and give the mix a new direction while it plays.

Explore dynamic soundscapes