Delibera

Inspiration

We kept asking: why do AI systems give confident answers to complex questions without ever deliberating? Humans don’t make high-stakes decisions by producing a single response; we explore alternatives, gather evidence, stress-test assumptions, and reconcile tradeoffs. Today’s multi-agent AI systems are either unstructured chat loops or rigid pipelines with no epistemic accountability. There’s no notion of claims, evidence, objections, or why a decision was made. Delibera was born from a simple conviction: decision-grade AI needs governance, not just generation.

What it does

Delibera is an open-source engine for structured, multi-agent deliberation. Given a question or decision, it:

Expands the problem into competing alternatives using a tree-structured search
Assigns specialized agent roles (proposer, researcher, red-teamer, validator, refiner) to each branch
Grounds claims in real evidence using Gemini + Vertex AI Search
Tracks an explicit epistemic ledger; every claim, evidence link, and objection is recorded
Scores, prunes, and converges branches using configurable, protocol-defined strategies
Produces a fully auditable decision artifact: final recommendation, supported claims with citations, resolved objections, and a complete trace

A Streamlit UI lets users explore the deliberation tree, inspect evidence, review objections, and understand why a particular option won.

Theoretical foundations

Delibera draws on several established research directions, but focuses on making them operational and auditable:

Tree-of-Thought and search-based reasoning

Instead of a single chain of thought, Delibera explicitly explores alternative branches, scores them, and prunes weak options through engine-governed search.

Argumentation theory

Reasoning is modeled as claims, supporting evidence, and objections, rather than as free-form text. Objections can block convergence unless resolved or explicitly accepted.

AI debate and adversarial roles

Dedicated red-team and validator agents are used to surface failure modes, challenge assumptions, and stress-test proposals.

Decision theory and bounded rationality

Deliberation proceeds under explicit budgets and stopping conditions. Convergence is deterministic and protocol-defined, not left to agent judgment.

Governance-first system design

Agents generate proposals, but never control the structure. Expansion, pruning, reduction, and convergence are enforced by the engine, producing traceable and replayable decisions.

The core contribution of Delibera is not a new model, but a governed deliberation architecture that integrates these ideas into a single, auditable decision process.

How we built it

Gemini API powering six distinct agent roles
Vertex AI Search grounding for evidence-backed claims (no hallucinated citations）
Declarative YAML protocols defining expansion rules, scoring, pruning, and convergence, with no hardcoded workflows
Streamlit UI with tree visualization (streamlit-agraph) and score plots (Plotly)
A multi-phase orchestration loop covering planning, expansion, validation, scoring, pruning, reduction, and finalization

Challenges we ran into

The sub-agents need to be more sophisticated to produce high-quality answers. It requires efforts beyond this hackathon.

Accomplishments that we're proud of

Full auditability, every decision can be reconstructed without re-running models
Evidence-grounded reasoning, claims must be supported, or they don’t survive
Governance by design, agents propose, but the engine controls the structure and convergence
Protocol-driven flexibility, change deliberation strategies with zero code changes