Tracer AI

Tracer AI Poster

Inspiration

We wanted a robust framework for full management of multi-agent

What it does

Multi-Agent Efficiency and Robustness:

Dynamic Agent LLM Selection: right-size the agent LLM based on agent role – prioritising smaller models. Dynamic Hallucination Protection: minimise hallucination with LLM as judge in the loop where needed.

Multi-Agent Transparency:

Detailed Visual Trace Map: map of agents, actions and tool calls with clear metrics per action and run. Interactive Explanation Agent: discuss the graph with an agent that can explain it in natural language using RAG.

Red-teaming:

Modified J2 Attack Agent: use chain-of-thought LLM to break refusal-trained LLM with iterative attacks.

How we built it

We built the multi-agent sustem using langchain, with models sourced from openrouter. We used search API from Valyu as well as many other open-source APIs. We used LangSmith API for observability.

Challenges we ran into

Learning new frameworks Versioning of modules API throttling Merge conflicts working together More ideas than time

Accomplishments that we're proud of

Novel approach to dynamic model selection for agents Working RAG for observable agent graph and nodes Predictive probabilistic approach to hallucination prevention