LinkedIn AI Agent

An automated agent powered by Playwright and LangChain to autonomously search for hiring posts, extract relevant profile details (authors, founders, senior engineers), and perform intelligent connections, reactions, and comments using an LLM.

Project Structure

• login.py: An interactive script that opens a browser for manual login and saves the authenticated cookies to state.json.
• browser_manager.py: A utility that re-uses the state.json file to launch headless/headed browsers for automated tasks, bypassing login walls.
• state.json (auto-generated): Contains local cookies for LinkedIn authentication.

Setup Instructions

1. Install Requirements:

   pip install playwright langchain langchain-openai langgraph pytest
   playwright install

2. Login Setup: Run the initial login script to generate your session state:

   python login.py

   Follow the terminal instructions, log in through the visible browser, and press Enter in the terminal to save your state.

(Further instructions mapping to Phase 2 and 3 will be deposited here as we build them out).

Progress

Phase 1: Foundation (Completed)

• MongoDB integration implemented using database.py with full Pydantic schemas.
• Playwright Stealth wrapper mapped inside browser_manager.py.
• Core Global Action Queue (action_queue.py) built to centralize & rate limit requests.
• System Health, Circuit Breaker, and Budget Managers built inside system_health.py and warmup.py.
• Slack integration commands /status and /pause mocked in slack_bot.py.
• Dry run parameters structured securely into schemas.

Phase 2: Connection Agent (Completed)

• LinkedIn people search scraper (people_scraper.py) to bypass layout limitations.
• Robust LLM integration (llm_scorer.py) utilizing LangChain and gpt-4o for relevance scoring and A/B template personalization notes.
• Agent Orchestration (agent_connection.py) parsing scrapes, verifying deductibles against db.connections, executing relevance checks, and pushing rate-limited connect actions into MongoDB queue.
• Acceptance Poller (acceptance_poller.py) scraping new connections and recording acceptances back into state.

Phase 3: Content Agent (Completed)

• Organic feed extraction scraper (content_scraper.py), supporting dynamic user sub-page target extractions.
• LLM interaction evaluation module (content_scorer.py) utilizing LangChain's structured output wrappers to evaluate posts on a 0-100 rubric, assessing high-quality viral properties or producing deep conversational comments based on post contexts.
• Digest dispatching functionality in slack_bot.py to organize morning notifications of potential reposts straight to Slack workspace channels for review.
• High-level orchestration wrapper (agent_content.py) mapping auto-reactions against the engage_list, scheduling likes and LLM-generated comments automatically without bypassing budget constraints.

Phase 4: Job Hunter Agent (Completed)

• Job extraction mechanism (job_scraper.py), configured to extract detailed job components directly from the dynamic search pane layout.
• AI relevance layer (job_scorer.py) leveraging LLMs to cross-match candidate profile keywords, scoring job fit objectively between 0-100 logic thresholds.
• Integrated #job-alerts Slack pipeline within slack_bot.py triggering instantaneous interactive job posts for roles scoring over a specific threshold.
• Action-queue bridging in agent_job_hunter.py queuing an organic "I've applied / sent my resume!" comment action directly below the poster's post if manually requested in the actual job descriptions.

Phase 5: Referral Agent (Completed)

• Direct employee extraction wrapper (referral_scraper.py) to bypass 1st-degree constraints and locate exact ICs (Individual Contributors) within targeted prospect companies.
• Context-aware LLM mapping model (referral_scorer.py) strictly structured via LangChain to draft authentic two-step Drip configurations: a no-ask non-sales connection request, and a humble ask-for-chat follow up depending on logic flow.
• The agent_referral.py execution orchestrator. This manages initializing wide net campaigns across Slack monitoring (#referral-campaigns) and schedules the Day-2 follow ups dynamically measuring 20+ hour delays precisely after your previous connections convert/accept through the global acceptance_poller.py. Drip interactions are pipelined directly to the central Action Queue buffer avoiding limit thresholds.

L i n k e d i n - M a r k e t i n g