Agora Capital

What Is Agora Capital?

Real investment committees have bulls, bears, risk officers, and compliance partners who structurally disagree with each other. A single AI analyst does none of that — it commits to its first answer, leaves no audit trail, silently suppresses minority concerns, and has no one challenging it.

Agora Capital applies the structure of a real investment committee to AI agents.

What Inspired It

Every VC partner we spoke to described the same problem: they see 500+ deals per year and can deeply diligence maybe 20. The other 480 get a 30-minute skim. Junior analysts produce shallow memos under time pressure. The fatal flaw in WeWork's unit economics, the related-party concentration at FTX, the technology validity gap at Theranos — these were all discoverable signals that a rigorous adversarial process would have caught earlier.

A single LLM given the same data produces one confident paragraph with no adversarial pressure, no audit trail, and no dissent. We wanted to build something that actually argues with itself before giving you an answer.

How We Built It

The Agent Committee

Nine specialized agents run in a 7-phase pipeline:

5 Specialist Agents (Financial, Market, Tech, Risk, Compliance) analyze the deal in parallel via asyncio.gather
BullCase builds the strongest possible case for the investment — it cannot concede without a counter-argument
BearCase builds the strongest possible case against — it cannot say "looks fine to me"
FactArbiter referees disputed claims through a 3-round evidence escalation protocol using real tool calls
DealLead orchestrates all phases and writes the final investment memo from verified claims only

Constitutional Governance

The committee operates under a 7-article charter with machine-enforced actions, not just guidelines:

Article 1 — every claim must cite a source; uncited claims are flagged automatically
Article 2 — if Bull and Bear agree in round 1, a mandatory adversarial round is triggered
Article 3 — risk score > 8/10 requires unanimous approval (not just majority)
Article 4 — a fatal quantitative metric (e.g. LTV/CAC < 1.0) forces a decisive PASS, immediately
Article 5 — a regulatory red flag pauses deliberation and routes to a human approval gate
Article 7 — aggregate confidence below 60% forces INVESTIGATE regardless of vote count

FactArbiter — Evidence Escalation

When BullCase and BearCase reach an unresolvable dispute, the FactArbiter doesn't pick a winner arbitrarily. It runs a 3-round protocol:

Round 1: Source quality evaluation — which citation is more recent, more credible, more specific?
Round 2: Tool call — a live web search, financial data lookup, or deterministic quant recompute. The tool result becomes the binding tiebreaker.
Round 3: If still unresolved, the claim is marked [DISPUTED] in the final memo with both positions preserved and committee confidence reduced by 15%.

Verdicts are earned through evidence — not assigned by consensus.

Deterministic Quant Gate

Before any agent speaks, every deal is scored against cited VC/SaaS benchmarks — LTV/CAC, gross margin, burn multiple, Rule of 40, client concentration (HHI). A diagnosed deal-killer immediately forces a PASS. A clean scorecard clears a decisive INVEST. The committee then reasons over computed numbers it cannot hallucinate.

Four Hallucination Defenses

Heterogeneous models — BearCase runs on a different model family (Llama 3.1 in local tier) so the adversary doesn't share Qwen's blind spots
Citation Verifier — upgrades Article 1 from "was a source named?" to "does the number match the input?"
Structured Abstention — agents can say "insufficient evidence," voting INVESTIGATE at ≤30% confidence rather than fabricating a position
Memory Write-Gate — agent reputation only credits grounded critical catches, not hallucinated ones

What We Learned

The hardest engineering problem wasn't the agents themselves — it was making disagreement meaningful. Early versions had agents that would argue briefly and then converge. We had to encode adversarial constraints at the system-prompt level AND at the orchestration level (Article 2 forcing an extra round when they agreed too quickly) to produce genuine tension.

We also learned that a deterministic quant layer changes everything about reliability. Once we added the arithmetic scorecard that pre-computes metrics before any LLM speaks, the committee's decisive accuracy jumped from ~0.2 to 0.9 on our benchmark. The agents are better at reasoning when the numbers they're reasoning over are locked.

The FactArbiter's tool-grounded dispute resolution — where a web search or quant recompute becomes the binding tiebreaker — is the single feature that most surprised us in testing. Watching the system call a tool mid-debate and flip a verdict based on the result felt qualitatively different from any other multi-agent system we'd seen.

Challenges We Faced

Scope discipline under a deadline. Nine agents across 7 phases with four hallucination defenses and a constitutional engine is a lot. We enforced strict token budgets (E-6 to E-9), capped disputes at 5 per session (E-7), and built a three-tier LLM architecture (mock → ollama → prod) so we could develop and test without burning cloud credits.
Benchmark integrity. Building a benchmark on famous failed companies (WeWork, FTX, Theranos) risks the model recognizing names and "knowing" the outcome. We built masked twin cases — identical financial numbers, renamed companies — to separate memorization from reasoning. The key result: Agora's decisive accuracy is unchanged on masked vs. named cases; the single-agent baseline degrades.
Human gate UX. When Article 5 fires mid-deliberation (regulatory red flag detected), the entire committee pauses and waits for human approval. Making that feel fluid in the frontend — with the React Flow graph visibly paused, a gate component appearing, and smooth resume after approval — required careful SSE state management.

Benchmark Results

We tested Agora Capital against a single Qwen agent on 10 historical deals (5 that failed catastrophically, 5 that became unicorns), using only information available at the time of investment and masked twins for contamination control.

Condition	Accuracy	Decisiveness	Fatal-Risk Detection
Agora — named companies	1.0	0.9	1.0
Agora — masked twins	1.0	0.9	1.0
Single agent — named	1.0	0.9	1.0
Single agent — masked	1.0	0.8	0.8

The provable edge is masking robustness: rename the companies and the single agent degrades (fatal-risk detection 1.0 → 0.8, decisiveness 0.9 → 0.8). Agora is unchanged because it reasons over computed financial metrics, not company names.