Agents

Prax keeps its main conversation loop lean by delegating domain-specific work to focused spoke agents. Each spoke runs its own LangGraph ReAct loop with a specialized system prompt and curated tool set.

Research shows that LLM tool-selection accuracy degrades past 20–30 tools (see Research). The hub-and-spoke pattern keeps the orchestrator’s tool count low while giving each spoke deep domain capabilities.

Contents

• Delegation — Hub-and-spoke delegation, spoke agents, sub-hubs, adding new spokes
• Self-Improvement — Fine-tuning pipeline, vLLM + Unsloth + LoRA hot-swap
• Self-Modification — Staging clone + verify + hot-swap / PR workflow
• Self-Regeneration — The recursive self-improvement loop (notice→propose→isolate→verify→canary→record) that drives the two surfaces above. Thesis: RSI is only as safe as its fitness function is un-gameable, so the eval-robustness stack (verify / auditor / accept-rate #22) is the precondition. Scope = harness self-improvement; graded-autonomy boundary; plugin-first first rung (#29)
• Checkpointing — LangGraph checkpoints with automatic retry