Skip to content

Concepts

Deep dives into the architecture and design of Mamba Agents.

Overview

  • Architecture Overview


    Understand how all the components fit together.

    Architecture

  • Agent Execution Loop


    How agents process messages and call tools.

    Execution Loop

  • :material-compress: Context Compaction Strategies


    Deep dive into the 5 compaction strategies.

    Compaction Strategies

  • Workflow Patterns


    ReAct, Plan-Execute, and other orchestration patterns.

    Workflow Patterns

Key Concepts

Agent as a Thin Wrapper

Mamba Agents wraps pydantic-ai with additional infrastructure:

┌─────────────────────────────────────────────────────┐
│                  Mamba Agents                      │
│  ┌─────────────────────────────────────────────┐    │
│  │              pydantic-ai Agent               │    │
│  └─────────────────────────────────────────────┘    │
│  + Context Management (message tracking, compaction)│
│  + Token Tracking (usage, cost estimation)          │
│  + Configuration (settings, env vars, files)        │
│  + Observability (logging, tracing, OTel)          │
│  + Error Handling (retry, circuit breaker)          │
└─────────────────────────────────────────────────────┘

Configuration Priority

Settings load from multiple sources:

1. Constructor arguments     (highest priority)
2. Environment variables     MAMBA_*
3. .env file                 project-specific
4. ~/mamba.env              user-wide defaults
5. config.toml / config.yaml file-based config
6. Default values            (lowest priority)

Built-in vs. Standalone Components

Most features work both integrated and standalone:

Feature Built-in (Agent) Standalone
Context agent.get_messages() ContextManager
Tokens agent.get_usage() TokenCounter, UsageTracker
Cost agent.get_cost() CostEstimator
Logging Automatic setup_logging()

Workflows vs. Agent Runs

Aspect Agent Run Workflow
Scope Single request Multi-step task
State Context maintained Full state tracking
Control Limited Fine-grained
Use case Chat, Q&A Research, analysis