Investigation Fleet

UI
investigation
architecture

The Problem

Technical regressions are easy to detect but often difficult to explain.

An API can suddenly become 10× slower. An ML model can suddenly lose accuracy. A deployment, preprocessing change, configuration change, or dependency may be responsible, but the symptom itself does not tell us which explanation is correct.

This creates a problem that is different from ordinary coding.

The engineer is not simply asking:

"Can you fix this code?"

They are asking:

"What caused this regression, what evidence supports that explanation, what alternatives have been investigated, and when do we have enough evidence to trust the conclusion?"

Modern AI assistants are already very good at individual tasks such as reading code, searching a repository, and running commands. Our focus is a different layer: automating the investigation process itself.

Our Approach

We built Investigation Fleet, an evidence-driven multi-agent investigation engine.

Instead of treating an incident as a single prompt, the system represents it as an investigation:

Incident
   ↓
Competing Hypotheses
   ↓
Experiment Selection
   ↓
Specialist Investigation
   ↓
Diagnostic Evidence
   ↓
Hypothesis Update
   ↓
Independent Verification
   ↓
Root Cause

#how fleet works
Fleet Commander
      │
      ├── Code Investigator
      │       └── Git / source analysis
      │
      ├── Runtime Investigator
      │       └── Controlled reproduction
      │
      ├── ML / Preprocessing Investigator
      │       └── Data and preprocessing analysis
      │
      └── Evidence Verifier
              └── Independent root-cause verification