Agent GlassBox

Inspiration

Modern AI agents make dozens of decisions, tool calls, and memory updates—but their reasoning is hidden. Debugging them feels like reading cryptic JSON dumps with no structure or narrative. Agent GlassBox solves this by turning raw execution logs into a transparent, human-understandable experience.

What it does

Agent GlassBox makes agent autonomy fully observable and explainable.

1. End-to-end Trajectory Capture

Tracks every step an agent takes: decisions, reasoning traces, memory updates, tool calls
Supports both single-agent ReAct-style flows and multi-agent workflows

2. Interactive Visual Inspector

Automatic graph visualization of reasoning → actions
Hierarchical task graph for multi-agent workflows
Temporal playback: step through the entire run like a debugger
Click any node to inspect arguments, outputs, timing, token usage

3. Built-in AI Analysis

Summaries of reasoning chains and decision paths
Automatic failure categorization
Bottleneck detection, execution stats, and improvement suggestions
Natural-language Q&A: “Why did step 0.3 fail?”

4. Three-Panel Workspace

Canvas: interactive graph with zoom/pan
Analysis Panel: auto-generated insights
Chat Panel: ask questions about the trace

Challenges we solved

Creating semantic layouts where spatial position reflects the cognitive process
Reconstructing workflow state from event streams
Rendering failed tool calls transparently instead of hiding them
Normalizing different log formats into one unified graph model