Post

CV-Pilot: A Hands-On Agentic Experiment

Posted Updated
Preview Image
By Gian Paolo Santopaolo
6 min read
CV-Pilot: A Hands-On Agentic Experiment

CV-Pilot: A Hands-On Agentic Experiment

I’m always on the lookout for ways to push the boundaries of what GenAI can do for productivity.

Here’s why I believe agentic AI is the next big leap, what its strengths and challenges are, and how everyone from startups to Salesforce is racing to build with it

Why Agentic AI?

Agentic AI frameworks—Microsoft AutoGen, Semantic Kernel, LangChain, CrewAI, and more (I’ve compared the most famous agenticAI framework in this blog post) —provide structured, goal-driven agents that coordinate to solve complex tasks. Unlike single-prompt chatbots, these systems:

  • Define specialized roles (data collector, analyzer, editor).
  • Chain together multi-step workflows with memory and tools.
  • Offer traceability and checkpoints for human oversight.
  • I wanted to try Crew AI on a real-world task, so I extended its resume-tailoring sample to match resumes to job offers and turned it into a fully working app.

Pros and Cons of Agentic AI

Pros

  • Autonomous Reasoning & Planning: Agents can self-check, plan sub-tasks, and adapt to dynamic goals, closing skill gaps in talent-scarce industries.
  • Efficiency & Focus: Routine, multi-step workflows—expense approvals, data extraction, onboarding—are automated, freeing humans for strategic decision-making.
  • Proactive Initiative: Modern agents anticipate needs, flag anomalies, or suggest next steps without explicit prompts).
  • Security Automation: In cybersecurity and support, agents can preempt incidents and handle high volumes of queries at scale.

Cons

  • Ethical & Bias Risks: Training data biases can propagate unfair decisions, and opaque decision paths complicate accountability.
  • High Costs & Complexity: Upfront investments in infrastructure, integration, and specialized talent can be prohibitive for smaller organizations.
  • Security & Privacy Vulnerabilities: Autonomous data access amplifies the attack surface, raising breach and misuse concerns.
  • Environmental Impact: Running large-scale, always-on agents consumes significant compute resources, heightening sustainability challenges.

Why Everything Is Going Agentic

The shift from static automation to dynamic, goal-oriented agents mirrors demands for adaptable, resilient systems. Agents don’t just execute code—they reason about objectives, recover from failures, and replan on the fly. As Harvard Business Review notes, we’re already seeing agents that can plan travel, manage supply chains, or act as virtual caregivers—ushering in a new era of human-AI collaboration. In finance and HR, AI agents are autonomously handling expense approvals and onboarding processes, boosting productivity without ballooning headcounts.

Big Companies Embracing Agentic AI

It’s not just startups riding the hype—Salesforce has rebranded its AI initiatives as “Agentforce,” deploying agents that manage over 80% of routine customer queries, cutting costs and improving satisfaction. Meanwhile, Andreessen Horowitz is pouring billions into agent-focused ventures, even as some warn of a marketing oversell. And under the hood, Microsoft and OpenAI continue to advance frameworks—AutoGen, Semantic Kernel, GPT-powered tools—cementing agentic architecture as the backbone of next-gen enterprise AI.

Looking Ahead: The Future of AI Agents

As foundational LLMs grow ever more capable, agents will not only execute tasks but learn and optimize over time, forming continuous improvement loops. The World Economic Forum predicts agents will shoulder specialized coding assignments, real-time analytics, and more—closing critical skill gaps in fast-moving sectors. With advances in model efficiency, multimodal reasoning, and on-device inference, the next wave of AI agents promises autonomy and adaptability far beyond today’s demos. Building off Crew AI’s flexible, role-based core, I’m excited to see how powerful agents will reshape workflows—and what new possibilities emerge as we hit higher LLM performance thresholds.

Why Crew AI Agentic Workflows?

Crew AI provides a flexible “multi‐agent” framework that lets you:

  1. Define agents with distinct roles and toolsets.
  2. Compose tasks that map to concrete, testable steps.
  3. Orchestrate everything under a manager agent with built-in planning, delegation, and traceability.

In a world where job applications are repetitive yet demand precision, this architecture seemed tailor-made to automate both CV customization and cover-letter drafting—without losing a human-in-the-loop checkpoint.

Introducing CV-Pilot

To test agentic by myself, I took the official Crew AI sample and built CV-Pilot—an AI-powered toolkit that automates resume tailoring and motivation-letter writing through two independent pipelines, gen_resume and gen_motivation. Here’s how I extended the sample, put its agentic orchestration to the test, and why the results are already looking very promising.

CV-Pilot consists of two CLI pipelines:

  • gen_resume

    • Input: raw cv_.md + job posting URL or Markdown
    • Output: docs/new_resume.md
    • Flow: extract requirements → profile match → tailored resume

  • gen_motivation

    • Input: new_resume.md + company_url + job posting
    • Output: docs/motivation_letter.md
    • Flow: scrape company site → parse job ad → draft intro/body/conclusion

Each pipeline dumps a JSON audit trail (application_state.json) so you can see every prompt, source URL, and intermediate draft. Human review tags (<!-- REVIEW: … -->) ensure you stay in control.

Extending the Sample: What I Changed

  1. Configuring Agents and Tasks I forked the Crew AI sample, then defined two new crews—JobApplicationCrew for resumes and MotivationLetterCrew for cover letters—each wired to my own config/agents.yaml and config/tasks.yaml. That let me specify different back-stories, goals, and tool sets per agent.

  2. Custom Tools

    • SerperDevTool for web search
    • ScrapeWebsiteTool for company info
    • DirectoryReadTool, FileReadTool, FileWriterTool for local docs All tools use simple caching to avoid redundant calls during development.

  3. Human-in-the-Loop After each draft pipeline, execution pauses. I review new_resume.md or motivation_letter.md, provide feedback in Markdown comments, then resume. This keeps the AI from “going off rails.”

  4. LangTrace Integration By plugging in the LangTrace SDK, I get spans around each agentic step, so I can visualize the end-to-end timeline in my observability dashboard.

Testing the Agentic Orchestration

With everything wired up, I ran a real job posting through both pipelines:

1
2
3
4
5
6
7
8
9
10
11

# 1. Tailor resume
python3 gen_resume.py --doc_path ./docs \
    --job_posting_url "https://jobs.example.com/123"

# (Review and adjust docs/new_resume.md)

# 2. Generate cover letter
python3 gen_motivation.py \
    --company_url "https://example.com" \
    --job_posting_url "https://jobs.example.com/123" \
    --doc_path ./docs

Key observations:

  • Relevance: The tailored resume highlights exactly the skills and keywords required in the posting.
  • Coherence: The motivation letter follows a clear intro–body–conclusion structure, citing mission statements and company values pulled directly from the website.
  • Speed: Both pipelines run in under a minute, even with multiple web scrapes.
  • Traceability: The JSON state file logs every decision—ideal for audits or fine-tuning prompts.

Results & Next Steps

I’m impressed by how seamlessly the agents collaborated:

  • The manager agent orchestrated extraction, profiling, and drafting without manual handoffs.
  • The editor agent produced polished output that only needed light human tweaks.
  • Switching the underlying LLM (OpenAI, Anthropic, Gemini) was as simple as changing an ENV var—no code changes.

Next on my roadmap:

  1. PDF/DOCX support: Let users drop in native resume formats.
  2. Parallel pipelines: Speed up large-scale tests by running multiple applications concurrently.
  3. Dashboard UI: Surface real-time progress and allow inline feedback in a web interface.
  4. Tone customization: Add sentiment analysis to adapt the cover letter’s style (e.g., formal vs. casual).

Looking Ahead: The Future of AI Agents

As foundational LLMs grow ever more capable, agents will not only execute tasks but learn and optimize over time, forming continuous improvement loops. The World Economic Forum predicts agents will shoulder specialized coding assignments, real-time analytics, and more—closing critical skill gaps in fast-moving sectors. With advances in model efficiency, multimodal reasoning, and on-device inference, the next wave of AI agents promises autonomy and adaptability far beyond today’s demos. Building off Crew AI’s flexible, role-based core, I’m excited to see how powerful agents will reshape workflows—and what new possibilities emerge as we hit higher LLM performance thresholds.

For more details, check out the full working sample on my GitHub repo

Agents, LLM, CrewAI
This post is licensed under CC BY 4.0 by the author.