zubbl-sdk 0.3.2

Self-Evolving AI Agent SDK - Make any AI agent 20-50% smarter automatically

Zubbl SDK

Self-Evolving AI Agent SDK — Make any AI agent smarter automatically.

Wrap your agent with one line. Zubbl records every execution, learns what works, and recommends better strategies next time. No code changes needed after initial setup.

Install

pip install zubbl-sdk

Quick Start

from zubbl import ZubblClient

zubbl = ZubblClient(api_key="zubbl_xxx")
smart_agent = zubbl.wrap(your_agent)

# Use exactly like before — now it learns and improves
result = smart_agent("Your task here")

That's it. Your agent now:

• Records every execution as a trajectory
• Extracts patterns from successes and failures
• Builds policies with actionable recommendations
• Injects learned strategies into future runs (~30 token overhead)
• Auto-recovers from failures using known fix patterns

How It Works

wrap(agent) → Agent runs → Record 3 steps automatically
                              ↓
                    Extract patterns by category
                              ↓
                    Build policy (confidence grows with data)
                              ↓
                    Next run: inject [Strategy: ...] prefix
                              ↓
                    Agent gets better — zero code changes

Benchmark Results

Metric	Result
Tasks to confidence 1.0	8 per category
Latency overhead	+85ms (+5%)
Actions learned per category	3 (plan → execute → validate)
Code changes required	Zero

Frameworks Supported

• OpenAI / Anthropic / Google Gemini
• LangChain
• CrewAI
• Any Python callable

Dashboard Feedback (No Code Changes)

1. Agent runs → trajectory recorded automatically
2. Go to app.zubbl.tech → rate the run → add feedback
3. Agent improves on next run — no code changes needed

Research Foundations

Built on peer-reviewed research:

• Training-Free GRPO (2025) — compact policy injection
• AgentHER (2026) — learn from failures via hindsight relabeling
• ETO (ACL 2024) — contrastive trajectory optimization
• Reflexion (NeurIPS 2023) — verbal self-reflection on failures

Links

• Documentation
• Dashboard
• Website
• Discord
• PyPI

License

MIT