Building an AI Board: Multi-Agent System Delivers Attributed Dissent

The FounderOS team built an AI board on Qwen, claiming two-cent, two-minute strategic memos. This multi-agent system identifies more risks and preserves dissent compared to a single LLM.

Solo founders and small operators frequently navigate critical, irreversible decisions without the benefit of a traditional board of advisors. The FounderOS team, comprised of Vincent Julijanto and Steven, addressed this by building an AI board of directors on Qwen. This multi-agent system is designed to generate structured debate and attributed dissent for strategic choices, claiming to deliver a full board memo for approximately two cents and two minutes per run. It offers a low-cost alternative to human advisors, specifically engineered to preserve diverse viewpoints.

Multi-Agent Architecture Preserves Dissent

The FounderOS system, detailed in its GitHub repository, operates as a LangGraph state machine. It orchestrates eight specialist agents, each with a defined role to ensure comprehensive analysis and the preservation of diverse viewpoints. A Scout frames the initial options, followed by Market Intelligence, which pulls cited benchmarks. Four parallel analysts (Trend, Finance, Growth, Capability) then run concurrently using asyncio.gather to provide specialized perspectives. The Skeptic receives all prior work and specifically attacks the weakest assumptions. A debate engine detects conflicts and runs rebuttal rounds. Finally, the Chair synthesizes the findings into a board memo. This architecture explicitly aims to prevent disagreements from being averaged out, instead reporting unresolved conflicts as attributed dissent.

Cost-Effective Decision Support

The FounderOS team reports that a full board run costs approximately two cents and takes about two minutes. This efficiency is attributed to a strategic split of tasks across Qwen models via the DashScope API. The faster qwen-turbo model handles operations like scouting, analysis, research, and memory indexing. The more capable qwen-plus model is reserved for heavier reasoning tasks, including the Skeptic's role, the debate engine itself, and the Chair's final synthesis. This tiered approach keeps the per-run cost low, making the "board" accessible for frequent, casual use.

Memory Without Embeddings

The system manages company-specific context through a markdown vault, akin to an Obsidian-style knowledge base. This approach stores a company profile, one note per decision, and includes dissent and outcomes. Crucially, the system foregoes vector embeddings for retrieval. Instead, qwen-turbo scans a small index of filenames, frontmatter, and one-line summaries to select relevant notes for the current question. The founders claim this method is easier to debug and more accurate for a single company's history compared to a vector store. The memory system demonstrated its efficacy by citing a previous decision's risk assessment and adjusting its recommendation in a subsequent, unrelated decision memo.

Benchmarking Against a Single Agent

The FounderOS team benchmarked their multi-agent board against a single qwen-plus call using an equivalent prompt. Across three live decisions, the board identified 9 distinct risks compared to 4 by the bare LLM. It also produced 7 instances of attributed dissent, while the single agent yielded 3. Raw outputs for this comparison are committed in the project's scripts/bench_results directory, allowing for audit of the counting methodology.

What We'd Change

The FounderOS approach offers a compelling blueprint for structured AI decision support, but several aspects warrant modification for broader application. The reported two-cent cost per run relies on the Qwen DashScope API. This specific LLM provider may not align with existing infrastructure or preferred models for many founders, particularly those deeply integrated with OpenAI or Anthropic. Replicating this cost-efficiency with other providers would require re-benchmarking and potential architectural adjustments.

The markdown vault memory system, while effective for debugging and a single company's history, presents scaling challenges. For companies with extensive, diverse knowledge bases or for a multi-tenant application, a simple index of filenames and summaries might become unwieldy. A more robust, potentially hybrid retrieval system incorporating embeddings for larger datasets would likely be necessary to maintain accuracy and performance.

The benchmark, while demonstrating a quantifiable improvement over a single-agent baseline, relies on a small sample of three decisions. Expanding this benchmark to include a wider variety of decision types, industries, and complexity levels would provide a more robust validation of the system's generalizability and the quality of the identified risks and dissent. The quality of "attributed dissent" is also subjective; further work would be needed to qualify its actionability.

Landing

The FounderOS project demonstrates that multi-agent AI systems can deliver structured, cost-effective decision support for solo founders. By explicitly designing for dissent and leveraging specialized agents, it moves beyond simple chatbot wrappers to provide a more nuanced analysis. The verifiable gains in risk identification and attributed dissent suggest a path for founders to access board-level strategic input without the traditional overhead.

The investor read

The FounderOS project highlights a growing trend in AI-driven decision support, particularly for the long tail of small businesses and solo founders underserved by traditional consulting or advisory boards. The low operational cost (two cents per run) is a critical factor, signaling potential for high-volume, low-margin SaaS models in this niche. While the current implementation uses Qwen, the underlying multi-agent architecture (LangGraph) is portable, suggesting investable opportunities in platforms that abstract away LLM dependencies or provide specialized agent libraries. The quantifiable improvement over single-agent LLMs (more distinct risks, attributed dissent) points to a market demand for structured, auditable AI outputs, moving beyond simple generative text. Investors should look for similar systems demonstrating robust benchmarking and clear value propositions in specific decision domains.