Skip to content
Back to Transmissions
// DATE
// AUTHOR
DHAATRIK
// CLEARANCE
PUBLIC
// INTERACTION

Why a Council of Models — Multi-Agent Process vs Single-Shot Certainty

Infinite Intelligence
====================================================================
// TRANSMISSION METADATA // QUICK REFERENCE (AEO/LLMO OBJECTS)
--------------------------------------------------------------------
- ENTITY: The Infinite Intelligence (github.com/dhaatrik/the-infinite-intelligence)
- VERSION: v4.0.0
- SCOPE: Client-side multi-agent council — parallel/sequential/round-robin, HITL, export
- MODEL: gemma-4-31b-it via @google/genai (BYOK)
- MOTIVATION: Single-shot LLM answers sound certain while skipping self-check
- KEY LESSON: Portfolio undersold v4 — four defaults + Beta squad, not "3 agents"
====================================================================

Mission Report: One Model Talking to Itself Is Still One Model

SYS.STATUS: TRANSMISSION LIVE // CLEARANCE: PUBLIC

Hey. My portfolio described three Gemini agents debating. v4.0.0 ships a four-agent default squad, optional Beta Mode dynamic personas, first-principles phase, critique rounds, human-in-the-loop steering, session branching, and PDF/Markdown executive export — all client-side with your own API key.

Not superintelligence. Process applied to models that otherwise skip peer review.


Mission Report: The Why — What Problem Was I Actually Solving?

Three frustrations stacked:

  1. Confident wrong answers. Single-shot completions sound authoritative when they are thin or biased.

  2. No adversarial pass. Creator without Critic is a monologue. I wanted structured roles: analyze, build, challenge, synthesize.

  3. Server rent on thinking. BYOK client-side execution keeps prompts and keys off my infrastructure — you pay Google directly, you own session data in the tab.

The Infinite Intelligence is a council-of-models workspace, not a chat wrapper.


Mission Report: What v4.0.0 Ships — Concrete Scope

CapabilityWhat it does
TopologiesParallel, sequential, round-robin debate (up to 5 critique rounds)
First principlesDeconstruct prompt into goals/constraints before agent work
Default squadFour specialized personas in DEFAULT_AGENTS
Beta ModeDynamically generated squad for the task
HITLApprove/reject/redirect agent outputs mid-run
BranchingFork conversation history at any node
ArtifactsSide panel for code/JSON extractions
ExportPDF (html2canvas + jsPDF) and Markdown reports
MetricsToken and cost estimates in UI

Model string per README: gemma-4-31b-it via @google/genai. Requires VITE_GEMINI_API_KEY.

Not shipped: Server persistence, automatic spend caps (you guard recursion depth).


Mission Report: Fuckups & Learnings

  • Undersold my own v4 on the portfolio. Three agents was stale; four defaults + Beta + HITL is the real story.
  • Unbounded debate loops cost money. Rate-limit bill fuckup is real — depth guardrails belong in product, not hope.
  • More agents can amplify noise if Critic prompts are loose. Role instructions matter more than headcount.
  • Debate is expensive in latency and tokens. UI must show progress, not fake thinking spinners.

Mission Report: Closing Transmission

Input a hard prompt. Watch the council argue. Steer it. Export the synthesis. Skepticism encouraged.

Tech: client orchestration stack. Costs: BYOK guardrails. Project page.