How sentient-engineering/jobber works
Project type and purpose
• Project type: A command-line AI agent framework in Python, with two implementations (“jobber” and “jobber_fsm”) that autonomously navigate the web and apply to jobs on behalf of the user.
• Purpose: Accept user tasks and preferences, plan a sequence of browser interactions, execute them via Playwright, extract or enter data, and iterate until the task (e.g., applying for a job) is complete.
File structure
Top-level:
• pyproject.toml, requirements.txt – dependency declarations
• .env.example – environment variables (OpenAI key, LangSmith key)
• README.md – project overview
• jobber/ and jobber_fsm/ – two parallel agent implementations
• test/ – end-to-end and unit tests
“jobber” and “jobber_fsm” packages both contain:
• __main__.py – CLI entrypoint
• config.py or config/ – config loading
• core/
– agents/ (planner_agent, browser_nav_agent)
– orchestrator (system_orchestrator or fsm/orchestrator)
– memory/ (LTM)
– prompts/ (prompt templates)
– skills/ (atomic browser interactions: click, open_url, screenshot, etc.)
– playwright_manager or web_driver/playwright – abstraction over Playwright
• user_preferences/user_preferences.txt – resume path, job filters
• utils/ – CLI helpers, DOM helpers, logger, Pydantic models, extraction utilities
README for additional insights
• Two implementations: “vanilla” conversational orchestrator vs. FSM-based orchestrator
• Dependencies on:
– OpenAI API (chat models)
– Playwright for browser control
– Optional LangSmith for tracing
• Setup steps emphasize:
– Launching Chrome in remote-debugging mode
– Providing .env keys
– Customizing user_preferences.txt
– Running via “python -m jobber” or “python -m jobber_fsm”
a. Main components
1. CLI / Entrypoint
2. Configuration loader (.env, config.py)
3. Orchestrator
– jobber: SystemOrchestrator
– jobber_fsm: FSMOrchestrator
4. Agents
– PlannerAgent (generates high-level plan via OpenAI)
– BrowserNavAgent (executes skill calls)
5. Memory store (LTM for contextual info)
6. Skills library (click, open_url, get_dom, extract_pdf, etc.)
7. Playwright Manager / WebDriver (controls Chrome)
8. External services
– Chrome instance (remote debugging)
– OpenAI API
– LangSmith (optional)
9. User preferences file (resume path, filters)
10. Logger & Tracing
b. Relationships & interactions (data flow)
• User runs CLI → loads config & preferences → starts orchestrator
• Orchestrator invokes PlannerAgent with task + memory → PlannerAgent calls OpenAI → returns plan
• Orchestrator dispatches plan steps to BrowserNavAgent → BrowserNavAgent calls skill functions → skill uses Playwright Manager → remote Chrome
• Skill results (DOM snapshots, screenshots, URLs) → returned to orchestrator → memory updated
• Loop until task completion
• Orchestrator logs events and (optionally) sends traces to LangSmith
c. Architectural patterns / design principles
• Modular design: clear separation of concerns (orchestrator vs. agents vs. skills vs. infrastructure)
• Two orchestration patterns: conversational vs. finite-state machine
• Dependency inversion: agents depend on abstract skill interfaces, not concrete Playwright calls
• Extensibility: add new skills or agents without changing core orchestrator
d. Key technologies/frameworks
• Python 3.x
• Playwright (via jobber/core/playwright_manager or jobber_fsm/core/web_driver/playwright.py)
• OpenAI API for LLM calls
• Pydantic for config validation
• LangSmith for optional tracing
• CLI utilities (argparse or click via utils/cli_helper)