OuroBoros

• 

  Main Page

• 

  After Critical error detected

• 

  Fixing Critical Error automatically

Inspiration

Autonomous agents are notoriously fragile. Most AI systems fail silently or enter infinite loops when they encounter complex reasoning traps, requiring manual intervention to reboot. We wanted to build a system that mimics a biological organism: a creature that not only survives its own failures but evolves immunities to them. The name Ouroboros represents the ancient symbol of a serpent eating its own tail, signifying the self-referential and regenerative nature of our AI loops.

What it does

Ouroboros is a resilient financial analysis agent system that self heals in real time. When the agent hits a "poison" prompt, a memory overflow, or a semantic loop, its "Nervous System" (Datadog) detects the anomaly. It then sends a distress signal via Kafka to a "Mechanic" agent. The Mechanic analyzes the failure, clears the corrupted context, and injects a "resurrection spell" that allows the agent to reboot with active immunity against that specific trap.

How we built it

The architecture is divided into three core components:

--> The Brain: A financial bot built using Vertex AI that performs multi step reasoning through planner, researcher, and analyst roles.

--> The Nervous System: Datadog APM and custom semantic analyzers that monitor "heartbeat" metrics and detect repetitive reasoning loops.

--> The Memory: A Confluent Kafka backbone that streams agent thoughts to a Next.js dashboard and handles remediation messages.

--> The Interface: A cyberpunk style SOC dashboard that provides live visualization of the agent's mental state and healing progress.

Challenges we ran into

Building a reliable "Semantic Loop Detector" was difficult. We had to implement cosine similarity checks to determine when an agent was repeating itself in spirit even if the exact words changed. Managing the state transition between a "Crashed" agent and a "Healed" agent required precise orchestration through Kafka to ensure the context was cleared without losing the original user intent. We also struggled with high latency in LLM tracing, which we optimized using custom decorators.

Accomplishments that we're proud of

We successfully demonstrated "Autonomous Resurrection." Seeing the agent hit a hard memory limit, crash, and then automatically restart and continue its task without human input was a major milestone. We are also proud of the "Threat Matrix" on our dashboard, which tracks the evolution of the agent's immunities as it encounters different types of technical and logic traps.

What we learned

We learned that observability is not just for humans. By making an agent’s internal metrics "observable" to another agent, we created a closed loop system far more resilient than a standalone LLM. We also gained deep experience in event driven architectures, specifically using Kafka to decouple the AI's "thoughts" from the UI's "display".

What's next for OuroBoros

The next step is to expand the "Mechanic" agent's capabilities so it can rewrite the agent's system prompt on the fly to avoid future traps. We also plan to implement "Multi Organism" orchestration, where multiple Ouroboros agents can share their learned immunities across a decentralized network, creating a collective immune system for AI.we ran into

Built With

• cloud
• css
• datadog
• ddtrace
• framer
• google
• kafka
• nextjs
• python
• react
• recharts
• tailwind