Skip to main content
HUMΛN
General
General

Trusted finished work in conversation

HUMΛN Engineering··5 min

OpenAI made voice feel natural. HUMΛN makes live AI work trustworthy.

Turn-based chat taught the world to prompt, wait, and hope. Live models listen while they speak, tolerate pauses, and recover from interruptions. That is table stakes for conversation — not differentiation.

Differentiation is what happens when the model hears "send it" and something actually sends — with delegation, risk classification, human approval when required, and a cryptographic receipt you can inspect later.


The chatbot trap

A charming voice that skips orchestration is worse than a boring form. It feels capable while bypassing the trust layer: no intent gate, no supersede when you change your mind mid-sentence, no proof of who approved what.

HUMΛN's Companion promise has always been: not a generic chatbot — your interface to the trust layer. LiveSession is how we ship that promise in product code.


One façade, three channels

LiveSession is the public API (/v1/live/sessions*). Under it, HumanOS runs three synchronized channels:

  1. Conversation — listen, speak, interrupt, repair (duplex transport is internal plumbing)
  2. Work — background analysis, connectors, agent delegation — without freezing the conversation
  3. Trust — Passport, delegation scope, approval gates, provenance stream

You experience one natural dialogue. HumanOS runs the machinery.


Interruptible by design (G1)

"Help me prep for my investor call — actually wait, start with what they'll challenge."

LiveSession keeps the session open. The first intent hypothesis is superseded — not silently overwritten in a prompt. No calendar mutation from a hypothesis you already abandoned. Provenance records the supersede so operators and auditors see the pivot.


Governed action (G3)

"Send the revised note to investor@example.com."

Intent may confirm, but Canon consequence policy still applies. High-risk sends surface approval.required before any connector fires. Say "don't send" and platform reflex (live.session.cancel_pending) cancels pending work — the agent does not get to negotiate.


Proof, not vibes

Every session carries a provenance_stream_id. Finished work links to receipts: who delegated, what policy fired, what executed. That is the HUMΛN moat — not voice quality alone.


Build with the same surface everywhere

You are Start here
Agent author ctx.live.openSession()SDK guide
App developer useLiveSession from @human/interaction-client/react
Operator human live startCLI reference
MCP client human.live.session.createintegration guide

Reference agent: live-session-coach maps golden scenarios G1–G3 to code.


What we are not shipping

  • Dual public APIs (live + duplex) — duplex is internal only at launch
  • STT push-to-talk as Companion voice UX — voice goes through LiveSession
  • Escape hatches (skip_approval, disable_overrides) — open API, closed loopholes

Try it

Open Companion and have a live text conversation that can revise, wait, and prove. Read the full architecture in our Live Interaction guide.

Intent → Live conversation → Governed action → Trusted finished work → Proof.

That is the product arc. LiveSession is the front door.