aragora

Self-Improving AI

Multi-agent debate with autonomous self-modification. Watch AI agents reason, critique, and evolve the system in real-time.

ar- (Latin: toward, enhanced) + agora (Greek: marketplace of ideas)

NOMIC LOOP ACTIVE

Watch Self-Improvement in Real-Time

See AI agents debate improvements to aragora itself. Every phase - from debate to design to implementation - is streamed live as it happens.

Open Live Dashboard

The Nomic Loop

aragora improves itself through structured cycles. Agents propose changes, debate their merits, implement them, and verify the results. The rules can change the rules.

1 Debate

Multiple agents propose and critique improvements. Consensus emerges from structured discussion.

2 Design

Winning proposals are refined into actionable specifications with clear success criteria.

3 Implement

Code changes are generated and applied. Each modification is tracked and reversible.

4 Verify

Tests run, linters check, and agents evaluate. Only verified changes proceed.

5 Commit

Approved changes are committed. Failed cycles trigger rollback to last known good state.

"In Nomic, the rules can change the rules. aragora applies this to AI - the system debates and implements its own evolution."

pip install aragora

Core Features

Heterogeneous Agents

Mix Claude, GPT, Gemini, Grok, Qwen, Deepseek, and local models. Different biases create stronger consensus.

Structured Debate

Propose, Critique, Revise loop. Configurable rounds and consensus mechanisms (majority, unanimous, judge).

Evidence Provenance

Cryptographic chain tracking sources. Every claim linked to evidence with reliability scoring.

Formal Verification

Z3-powered proof checking. Verify logical claims with SMT solver integration.

Decision-to-PR Pipeline

Turn debate outcomes into GitHub PRs. Risk registers, test plans, implementation specs.

Red-Team Mode

Adversarial testing with steelman/strawman attacks. Find weaknesses before production.

Why Multi-Agent Debate?

Single Model aragora (Multi-Agent)
One perspective Heterogeneous viewpoints
Hallucinations go unchallenged Agents critique each other
Black box reasoning Transparent debate transcript
Single point of failure Consensus requires agreement
No dissent recorded Minority views preserved
$ aragora ask "Design a rate limiter for 1M req/sec" [Round 1] claude_proposer: Token bucket with Redis cluster... [Round 1] gemini_critic: Race condition in distributed counter [Round 1] gpt_critic: Missing backpressure mechanism [Round 2] claude_proposer: Revised with CAS operations... [Round 2] gemini_critic: Addresses race condition [Round 2] gpt_critic: Added circuit breaker Consensus reached (confidence: 87%) $ aragora nomic --cycles 3 # Run self-improvement loop [Cycle 1] Debating improvements... [Cycle 1] Implementing: "Add caching to debate storage" [Cycle 1] Verified ✓ Committed

🚀 Use Cases

🔍

Code Review

Multi-agent security & quality analysis

🛠

System Design

Debate architectural decisions

🔥

Incident Response

Rapid multi-perspective RCA

📚

Research Synthesis

Combine findings, challenge claims

🔒

Security Testing

Red-team your proposals

Decision Making

Scenario analysis with dissent tracking

Philosophical Foundations

"Truth emerges from the marketplace of ideas, not central authority."

Voluntary Exchange

Agents participate freely. Consensus emerges from debate.

🔗

Counter to Monolithic AI

Alternative to "trust the one big model."

🌏

Decentralized Coordination

No single authority decides truth.

💡

Emergent Order

Better answers emerge from competition.

Technical Inspiration

Stanford Generative Agents LLM Multi-Agent Debate ChatArena Project Sid Nomic (Peter Suber) Agorism (SEK3)
Watch Live View on GitHub PyPI Package