v0.1.0 — Initial Release

Blitz-Swarm v0.1.0

A parallel multi-agent swarm architecture for collective AI intelligence.

All agents live. All agents share memory. All agents iterate until consensus.

What's Inside

• Parallel agent invocation — all agents fire simultaneously via asyncio.gather(), no sequential waves
• 5 specialized roles — Researcher, Critic, Fact-Checker, Quality Judge, Synthesizer
• LLM-powered planning — dynamic agent count and subtopic assignment based on topic complexity
• Iterative consensus — swarm iterates until unanimous agreement, with holdout override after round 3
• Dissent preservation — minority views are first-class output, never suppressed
• G-Memory — three-tier hierarchical memory (NeurIPS 2025) that makes the swarm smarter with every task:
  • Tier 1: Interaction graphs (raw agent traces)
  • Tier 2: Query graph (semantic task-level links)
  • Tier 3: Insight graph (LLM-distilled cross-task wisdom)
• Redis blackboard — live agent coordination with append-only streams (optional — runs fine without)
• SQLite+WAL + LanceDB — persistent storage with single-writer serialization
• Memory eviction — archival and purging to prevent unbounded growth
• TOML config — all knobs tunable via blitz.toml

Quickstart

# Run without Redis
python orchestrator.py "SQLite WAL mode internals" --no-redis

# Preview agent plan
python orchestrator.py "CRDT conflict resolution" --dry-run

Stack

Python 3.12+ · Claude CLI · Redis (optional) · SQLite · LanceDB · sentence-transformers (MiniLM)

Philosophy

No frameworks. No magic. Vanilla Python + asyncio + subprocess. Every component is explicit and inspectable. Model-agnostic — any CLI-invocable model works. Privacy-first — all memory is local.

---

MIT License