How It Works¶

Zubbl uses a dual learning loop to continuously improve your AI agents.

The Dual Loop¶

graph TB
    subgraph "Loop 1: Trajectory Learning"
        A[Agent executes task] --> B[SDK records trajectory]
        B --> C[Patterns extracted & scored]
        C --> D[Policies updated]
        D --> E[Better recommendations next time]
    end

    subgraph "Loop 2: Reflexion"
        F[Agent fails] --> G[LLM analyzes failure]
        G --> H[Generates fix strategy]
        H --> I[Retries with insight]
        I --> J{Success?}
        J -->|Yes| K[Learn recovery pattern]
        J -->|No| L[Escalate to human]
    end

    E --> A
    K --> C

Loop 1 — Trajectory Learning¶

Every time your agent runs a task, the SDK records the full trajectory: what steps it took, which tools it used, how long it took, and whether it succeeded. These trajectories are analyzed to extract patterns — what works for which types of tasks. Over time, when you call query(), you get recommendations based on what has historically worked best.

Loop 2 — Reflexion¶

When an agent fails, instead of just logging the error, the SDK triggers an LLM-powered analysis of what went wrong. It generates a fix strategy and retries the task with that insight in-context. ~33% of failures are recovered on the first retry.

SDK Integration¶

sequenceDiagram
    participant App as Your App
    participant SDK as Zubbl SDK
    participant API as Zubbl API

    App->>SDK: start_trajectory("code review")
    SDK->>SDK: Begin recording

    App->>SDK: record_step("scan", "semgrep", ...)
    App->>SDK: record_step("fix", "code_gen", ...)

    App->>SDK: end_trajectory(SUCCESS)
    SDK->>API: POST /ingest (trajectory)
    API-->>SDK: accepted + queued for learning

    Note over API: Pattern extraction runs async

    App->>SDK: query("code_review")
    SDK->>API: GET /query?task=code_review
    API-->>SDK: Recommended strategies
    SDK-->>App: Use these patterns

What Gets Stored¶

Data	Purpose	Retention
Trajectories	Steps, tools, outcomes	90 days
Patterns	Extracted strategies with scores	Permanent
Policies	Recommended actions per task type	Updated continuously
Feedback	Human/automated ratings	Permanent

All data is tenant-isolated — your data is never shared with other tenants or used to train models for other users.

Security¶

Encryption: All data encrypted in transit (TLS 1.3) and at rest
Authentication: API key-based authentication
Rate Limiting: Per-tenant rate limiting on all endpoints
Tenant Isolation: Full data isolation across all tenants