Sovara¶
The trace of an agent executing a task is effectively a DAG, where the nodes are LLM or tool calls, and the edges are virtual connections representing data that flows from one node to another. However, current AgentOps platforms do not represent agentic traces like this, making it hard to understand where and why an agent failed to complete a task.
Sovara is a development tool that creates interactive dataflow graphs of agent traces, enabling visualization, editing, and debugging of data flow in agentic systems – with zero code changes.
Overview¶
Sovara goes beyond being a simple observability tool:
- Visualize agent traces as a DAG - See how data flows between LLM and tool calls in your application
- Edit inputs and outputs - Modify LLM and tool call inputs/outputs and re-run with changes, where previous nodes in the DAG are cached
- Debug dataflow - Track how LLM outputs propagate through your code
- Automatically improve any agent - Sovara is also an MCP tool, enabling observability and fast debugging for your favorite coding agent like Claude Code
How to use¶
We assume you have coded your workflow in Python, i.e., you run it like this:
All you change is the Python command. Whenever you want to develop with Sovara, run:
You can set a custom run name using --run-name:
This feels exactly the same as running Python but also analyzes your code and populates our VS Code extension:
- Program prints/reads to/from the same terminal, crashes the same, etc.
- You can use the VS Code debugger normally
For running evaluations or batch processing, use the launch context manager to create separate traces for each sample:
from sovara import launch
for sample in samples:
with launch(f"sample-{sample.id}"):
result = evaluate(sample)