4.1.5. Agent Internals

Agent Container Contents

The agent Docker image (services/agent/Dockerfile) is based on python:3.12-slim and includes:

The container runs as a non-root user (celeryuser, UID 1000).

Startup Sequence

entrypoint.sh
  │
  ├─ [optional] pip install requirements.txt (base64 decoded)
  │
  └─ exec celery -A main worker
       --queues=$RUN_ID
       --concurrency=1
       -n worker$POD_INDEX
       --prefetch-multiplier=1
            │
            ├─ Python imports main.py
            │    ├─ load_agent_message() from Redis
            │    ├─ Create Celery app
            │    └─ exec(agent_message["code"], {app, agent_message})  [details]
            │         └─ Project's main.py now loaded
            │
            └─ Celery signal: worker_process_init
                 └─ init_worker()
                      ├─ asyncio.new_event_loop()
                      └─ AgentContext.__aenter__()
                           ├─ ProxyManager.launch_proxy()
                           ├─ FingerprintManager (ready)
                           ├─ DatabaseConnector (connected)
                           ├─ Screen.launch_screen()
                           └─ Browser.launch()

For how the bundled code reaches the agent via Redis, see Code Bundling and Execution. For how the agent message is constructed from config layers, see Configuration Pipeline.

Concurrency model

Each agent pod runs exactly one Celery worker process with concurrency=1 and prefetch-multiplier=1. This means:

• One job at a time per pod
• One browser instance per pod
• No contention on the local mitmproxy (port 8081)
• Predictable resource usage

Available Modules

Since your project code executes inside the agent container, it has direct access to several built-in modules through the AgentContext and JobContext objects. These modules are initialized automatically during the agent lifecycle (see Startup Sequence and Context Managers above) and are ready to use in your task code.

Each module has a dedicated guide covering configuration and usage:

• ProxyManager — Local mitmproxy layer for traffic interception, upstream proxy rotation, and addon system. See the Proxy Provider guide.
• FingerprintManager — Fetches forged browser fingerprints from the Fingerprint Provider service. See the Fingerprint Provider guide.
• DatabaseConnector — SQL Server connection wrapper for reading and writing data. See the Database Connector guide.
• MonitoringClient — Reports run/job lifecycle events to the Backend API for real-time dashboard updates. See the Monitoring Client guide.