Zubbl SDK

Self-Evolving AI Agent SDK — Wrap any AI agent. It learns, recovers from failures, and gets better over time.

Install

pip install zubbl-sdk

Quick Start (3 Lines)

from zubbl import ZubblClient

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

result = agent("Review this PR for security issues")

That's it. Your agent now:

1. Auto-recovers from failures — retries with fixes, no crash
2. Learns from every run — records trajectories, extracts patterns
3. Injects proven strategies — prepends natural-language guidance before each task
4. Gets better over time — policies improve as data accumulates

How It Works

wrap(agent) → Agent runs → Trajectory recorded
                              ↓
                     Patterns extracted by task type
                              ↓
                     Policy built (confidence grows with data)
                              ↓
                     Next run: natural-language guidance prepended
                              ↓
                     Agent performs better — zero code changes

Framework Examples

OpenAI

import openai
from zubbl import ZubblClient

client = openai.OpenAI()
zubbl = ZubblClient(api_key="zubbl_xxx")

def my_agent(task):
    return client.chat.completions.create(
        model="gpt-4o",
        messages=[{"role": "user", "content": task}],
    ).choices[0].message.content

agent = zubbl.wrap(my_agent)
result = agent("Write unit tests for auth.py")

LangChain

from langchain_openai import ChatOpenAI
from zubbl import ZubblClient

llm = ChatOpenAI(model="gpt-4o")
zubbl = ZubblClient(api_key="zubbl_xxx")

agent = zubbl.wrap(llm.invoke)
result = agent("Explain this error traceback")

Any Callable

zubbl = ZubblClient(api_key="zubbl_xxx")
agent = zubbl.wrap(any_function_that_takes_a_task)

Auto-Recovery

When your agent fails, Zubbl catches the error and retries with a fix:

agent = zubbl.wrap(my_agent)  # auto_recover=True by default

# If my_agent crashes internally:
# 1. Zubbl catches the error
# 2. Applies a fix (prompt patch, retry, model swap)
# 3. Returns the correct result
# No crash. No code change.

Custom Policy Hook

Control exactly how learned strategies are injected:

def my_hook(args, policy):
    task = args[0]
    tips = "\n".join(policy.recommended_actions[:3])
    return (f"Tips from past runs:\n{tips}\n\nTask: {task}",) + args[1:]

agent = zubbl.wrap(my_agent, policy_hook=my_hook)

Dashboard

1. Agent runs are recorded automatically
2. Visit app.zubbl.in to view trajectories and rate runs
3. Your feedback improves policies — 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