MO§E§™ — Three-Scenario Execution Comparison (v7)

MO§E§™ — Three-Scenario Execution Comparison · v7

Current System vs New System:
Same Input. Same Agents. Three Architectures.

Commit state before intelligence branches.

All three scenarios run in the same session against the same utterance. Scenario 1 simulates the current cascaded system — agents receive each other's text output, full NLU re-inference, handoff overhead applied. Scenarios 2 and 3 run the K-governed architecture. The comparison is honest: same agents, same task, different architecture.

§ Anthropic API Key — click to expand cached results mode

Enter your Anthropic API key to run the demo live. The key is stored only in this browser session (sessionStorage) — it is never persisted to disk, transmitted to any server other than Anthropic, or visible to anyone else. Without a key, the demo replays a cached run so you can still see the output.

Scenario 1 — Current System

Cascaded text relay

Representative cascaded orchestration (text-dominant; partial structure not assumed). Full NLU inference. Each agent receives the prior agent's text output. Sequential by necessity. Handoff overhead at every step. No shared kernel.

Scenario 2 — New / Sequential

K-governed, sequential scheduler

Commitment kernel extracted once. All agents receive K (JSON). Sequential scheduler. No re-inference. No text relay between agents.

Scenario 3 — New / Parallel

K-governed, near-parallel scheduler (staggered launch)

Same K. Same agents. Parallel execution — all agents access K simultaneously. Wall-clock ≈ slowest single agent.

Caller Utterance Input for all three scenarios

S1 — Current System: Cascaded Text Relay

—

Full NLU inference → agents receive each other's text output → handoff overhead (500ms) between each agent → sequential by necessity.

NLU / Intent

input: raw text

—

Waiting...

Tone Agent

input: NLU text

—

Waiting...

Subject Agent

input: tone text

—

Waiting...

Response Agent

input: subject text

—

Waiting...

S2 — New System: K-Governed Sequential

—

K extracted once. All agents receive JSON kernel. Sequential scheduler. No re-inference. No text relay.

Tone Agent

input: K

—

Waiting...

Subject Agent

input: K

—

Waiting...

Response Agent

input: K

—

Waiting...

Pace Agent

input: K

—

Waiting...

S3 — New System: K-Governed Parallel

—

Same K. All agents access simultaneously (staggered 200ms to avoid rate limit). Wall-clock ≈ slowest single agent.

Tone Agent

input: K

—

Waiting...

Subject Agent

input: K

—

Waiting...

Response Agent

input: K

—

Waiting...

Pace Agent

input: K

—

Waiting...

Copied to clipboard