Structured argumentation. Fault-tolerant consensus. Cryptographic auditability.
Part of the [&] Protocol stack · Ecosystem overview · Three-protocol stack · Stack status
Welcome to the documentation hub for Deliberatic — an open-source deliberation protocol for multi-agent AI systems. Deliberatic provides structured argumentation, Byzantine fault-tolerant consensus, and Merkle-chained evidence logs for agent decision-making.
The protocol extends Dung's Abstract Argumentation Framework (1995) into a weighted bipolar system with agent authorship, integrates a two-phase consensus mechanism (Raft fast-path / PBFT conflict-path), and produces cryptographic evidence chains for regulatory compliance.
Multi-Agent Debate (MAD) research has demonstrated that LLM agents improve decision quality through structured disagreement. But current MAD frameworks suffer three critical gaps:
No formal argumentation semantics — agents exchange opinions with no
structured way to compare, attack, or support positions
No Byzantine fault tolerance — a compromised agent can corrupt the
entire cluster
No audit trail — enterprise deployments require explainable, verifiable
decision chains
Deliberatic addresses all three.
Homepages
Root Docs
Deliberatic Docs
A formal tuple DAF = <A, R-, R+, w, a, C, p> combining:
Weighted bipolar argumentation (attacks + supports with evidence-derived weights)
Agent authorship tracking (every argument traces to its submitting agent)
Constitutional constraints (hard boundaries + soft preferences)
Domain-aware reputation (ELO-derived, calibration-adjusted)
Fast path (Raft-derived) — when positions have a clear winner (~50ms)
Conflict path (PBFT-derived) — when positions are close, Byzantine-tolerant (~200ms)
YAML-based normative constraints: hard boundaries (instant rejection on violation), soft preferences (score adjustments), and amendment rules (vindicated dissents trigger constitutional votes).
Every deliberation round produces an immutable evidence chain — SHA-256 hashed positions, challenges, votes, and verdicts with a verifiable Merkle root.
| Protocol | Support |
|---|---|
| A2A (Google/Linux Foundation) | Native — agents advertise deliberation/v1 skill |
| MCP (Anthropic/AAIF) | MCP server with 7 tools |
| HTTP | JSON-RPC 2.0 + SSE for streaming |
| Tool | Description |
|---|---|
deliberatic_open_round | Open a new deliberation round |
deliberatic_submit_position | Submit a position with evidence |
deliberatic_challenge | Challenge another agent's position |
deliberatic_resolve | Trigger resolution and get verdict |
deliberatic_query_evidence | Query evidence chains by round ID |
deliberatic_get_reputation | Get agent reputation scores |
deliberatic_amend_constitution | Propose constitutional amendment |
Deliberatic is the decision layer:
| Layer | Product | Role |
|---|---|---|
| Distribution | FleetPrompt | Agent runtime, skill execution |
| Standards | SpecPrompt | Execution specs, schema validation |
| Orchestration | Delegatic | Task routing, workflow graphs |
| Decisions | Deliberatic | Argumentation, consensus, audit |
| Execution | AgenTroMatic | Autonomous task execution |
| Infrastructure | WebHost.Systems | Hosting, deployment |
| Spatial | GeoFleetic | Location-aware routing |
| Temporal | TickTickClock | Scheduling, deadlines |
Spec: Protocol Specification
[&] Protocol ecosystem: AmpersandBoxDesign/
[&] Ampersand Box Design — deliberatic.com