Skip to main content

The Project

ISEE v2 (Idea Synthesis and Extraction Engine) is a multi-agent system I built in 2026 to solve a specific problem: how do you systematically generate a large set of diverse LLM perspectives from a single query and reliably extract the best ones? This is a complete reconceptualization of ISEE v1, which I built in 2025. I learned so much building v1, and in the year since, I got clearer about what I wanted it to do and what I wanted the user experience to be. I also developed a sharper understanding of agentic engineering patterns. v2 isn’t an iteration; it’s a ground-up rebuild using a more precise and methodical development process. As a result, ISEE v2 orchestrates 6 specialized agents through a fixed pipeline, generates ~60 responses through multi-model synthesis, clusters them by intellectual angle, then runs a structured tournament debate where advocates argue and a skeptic challenges until only the strongest three ideas survive. The user can read the briefing only and/or a transcript of the entire debate.

What I Built

1. Multi-Agent Pipeline

Six specialized agents, each with a distinct role and strict input/output contracts:
AgentRole
Prep AgentGenerates 3-5 knowledge domains specific to the query
Clustering AgentGroups ~60 responses by intellectual angle (not source)
Advocate AgentsEach argues for their cluster’s value
Skeptic AgentChallenges all advocates with specific critiques
Synthesis AgentSelects 3 ideas, writes the briefing
Translation AgentConverts to plain language with action items

2. Multi-Model Synthesis Layer

Runs the query through 6 heterogeneous LLMs × 11 cognitive frameworks × 3-5 dynamically generated domains — roughly 60 parallel calls. The models are chosen for cognitive diversity.

3. Tournament Debate System

Clustered ideas compete in a structured debate where Advocates argue, a Skeptic challenges each argument, and Advocates either rebut or concede. Low-quality responses self-eliminate when they can’t defend against skeptical challenges. The user sees which ideas survived and why.

4. Operations Dashboard

Tracks how much each run costs, which models are reliable, and where time is being spent. Shows cost breakdown by provider and model, success rates for each LLM, and how long each pipeline stage takes. Result: github.com/joseph-fajen/isee-v2

6 Specialized Agents

Plan-and-execute pipeline with bounded autonomy. Each agent has strict TypeScript contracts.

~60 Parallel LLM Calls

6 LLMs × 11 frameworks × 3-5 dynamic domains. Heterogeneous models for genuine cognitive diversity.

Tournament Debate

Advocates argue, skeptic challenges, rebuttals defend. Ideas that can’t survive scrutiny are eliminated.

Open Source

Full codebase available on GitHub.

Tech Stack

TypeScript · Bun · Anthropic Claude SDK · OpenRouter · Multi-Agent Orchestration · SQLite