Skip to content
Back to Project Archive

Infinite Intelligence

Multiple Gemini agents debate a problem before you get a final answer — not one model talking to itself in a mirror.

Infinite Intelligence Logo
// PROJECT_ID
the-infinite-intelligence
// AUTHOR
DHAATRIK CHOWDHURY
// SUBJECT
CASE STUDY
// CLASSIFICATION
PUBLIC // OPEN
// TELEMETRY
STATUS: OPERATIONAL // AGENTS: DEBATING // PARALLEL: TRUE
// IMPACT.LOG
// FUCKUP_LOG
RECORDED
[ FUCKUP LOG ]  I kicked off a multi-agent debate test loop without a maximum recursion depth block, waking up to a rate-limited API key and a massive bill.

SYS.STATUS: Council-of-agents pipeline works — Creator, Critic, Synthesizer all shipping

One model giving you a confident wrong answer is still one model. The Infinite Intelligence is my experiment in making AI argue with itself before you trust the output.

Why I started this

I kept hitting the same wall: single-shot LLM responses sound authoritative even when they’re thin or biased. I wanted a workflow where specialized roles — generate, critique, refine — run in a structured loop, so the final answer survived at least one round of peer review.

Not magic. Not “superintelligence.” Just process applied to models that otherwise skip self-checking.

What I tried (and what broke)

I modeled the system as a state machine. Agents (Creator, Critic, Synthesizer) pass messages through typed transitions in TypeScript 5.8. Google Gemini handles the roleplay via system instructions — different personas, different jobs.

Three workflow modes shipped from the build:

  • Parallel collaboration — agents work concurrently where independence helps
  • Sequential workflows — strict ordering when each step depends on the last
  • Round-robin debate — iterative critique until synthesis

React 19 + Vite 6 render the interaction canvas — watching agents talk in real time is half the point. Getting the UI to stay smooth while multiple async completions land out of order took careful state design.

Fuckups & learnings

  • More agents ≠ better answers. Extra rounds can amplify noise if the Critic isn’t constrained. Role prompts matter more than agent count.
  • State machines pay off fast. Ad-hoc promise chains became unmaintainable by week two.
  • Debate is expensive. Latency and token cost are real; the UI should show progress, not a spinner pretending it’s thinking.

Where it stands now

All three workflow modes are implemented. You input a prompt and watch the council debate, critique, and refine live before the synthesized output lands.

Closing transmission

If you’re skeptical of single-model certainty — good, me too. Clone it, break a workflow, and see if the council catches what one shot missed.