Mini.Us

Inspiration

The inspiration for Mini.Us comes from the concept of the "Drawin Golden Machine": a system capable of self-replication and self-improvement. We wanted to move beyond static, single-purpose AI agents and build a truly dynamic ecosystem where an agent isn't just a tool User uses, but agents that can build the right tool for any task. The dream was to have a "Mini Me" that grows alongside you, spawning specialized workers to handle complexity so you don't have to.

What it does

Mini.Us is an advanced multi-agent orchestration platform that implements the agent spawning and meta agent pattern.

Dynamic Agent Spawning: It analyzes your complex requests and dynamically generates specialized sub-agents (Python code) to handle specific parts of the task. Secure Code Execution: It safely runs potentially risky generated code and data analysis tasks in isolated Daytona sandboxes. Full Observability: It provides deep visibility into the "mind" of the agent mesh using Sentry, allowing you to trace the decision-making process across multiple autonomous agents. Planning & Execution: It breaks down high-level goals into actionable plans and delegates them to its dynamically created workforce.

How we built it

We built Mini.Us using a modern, component-based AI stack:

Core Framework: Built on the Strands Agents SDK, utilizing its Agent and Tool abstractions for a modular architecture. Sandboxing Engine: We integrated Daytona to create ephemeral, secure execution environments. When the main agent decides it needs a new capability, it writes code that is executed inside a Daytona sandbox to configure and instantiate a new sub-agent. Observability Layer: We embedded Sentry tracing throughout the orchestration layer. Every agent thought, tool call, and sub-agent delegation is captured as a span, creating a complete distributed trace of the AI's workflow. Frontend: A React and Streamlit (for backend test) interface provides dynamic real-time interaction, allowing users to chat with agents, visualize plans, and see artifacts generated by the sandboxed environments. Models: We leveraged OpenAI's GPT family for high-reasoning capabilities required for code generation and planning.

Challenges we ran into

The "Infinite Loop" Risk: When agents create agents, there's a risk of runaway recursion. Debugging this was incredibly difficult until we integrated Sentry. The visualization of the trace tree was a game-changer for identifying where logic loops were happening. Sandboxing Complexity: Safely executing generated code is hard. We initially struggled with dependency management in ephemeral environments. Daytona solved this by giving us fully isolated workspaces where we could pre-install requirements and destroy the environment immediately after use. Context Management: Passing state between the main agent and its dynamically created children was tricky. We had to design a robust protocol for context propagation to ensure sub-agents understood the global goal.

Accomplishments that we're proud of

The "Golden Machine" Reality: We successfully demonstrated an agent writing code to create another agent, which then successfully executed a task. Seeing that recursive creation work for the first time was magical. Seamless Daytona Integration: We built a smooth pipeline where code moves from the agent's "mind" to a Daytona sandbox and back as a structured result in seconds. Crystal Clear Observability: Turning an opaque multi-agent process into a clear, navigable Sentry trace felt like turning on the lights in a dark room.

What we learned

Observability is not optional: In multi-agent systems, you cannot debug with print statements. Distributed tracing is essential to understand emergent behavior. Ephemeral is better: For AI-generated code, spinning up a fresh environment (Daytona sandbox) for every execution prevents side-effects and pollution, making the agents much more reliable. Agents need structure: Giving agents the ability to code is powerful, but they need strong architectural patterns (like "Agents-as-Tools") to be effective rather than chaotic.

What's next for Mini.Us

Our vision for Mini.Us is to create a full working General Purpose AI agent tha uses meta agent that creates agent based on new agent.

Built With

• daytona
• sentry