AI job postings grew 117% year-over-year. "Agentic engineer" barely existed as a job category 24 months ago. Today it's everywhere — on resumes, in recruiter pitches, and in company hiring pages for roles that don't quite know what they're looking for.
The result is a talent market flooded with candidates who've completed a LangChain tutorial, deployed a RAG pipeline on a weekend, and are now calling themselves agentic engineers. That's not malicious — it's just what happens when a new category emerges faster than the industry can develop standards for it.
Most staffing firms can't navigate this. Their screeners were trained on a different era's criteria. Here's how Minimalistech actually vets for agentic engineering capability — and why the difference matters.
What Makes Agentic Engineering Actually Different
Before you can vet for something, you have to understand what it is. "Agentic engineer" is not a fancy synonym for "developer who uses AI tools." It describes a specific technical discipline with its own architecture patterns, failure modes, and mental models.
The core skills that define a genuine agentic engineer:
- Multi-agent orchestration: Designing systems where multiple AI agents collaborate, hand off work, and maintain coherent state across sessions. This is a distributed systems problem with an LLM twist.
- Tool-calling architecture: Building reliable tool interfaces that agents can invoke — handling retry logic, partial failures, and unexpected outputs without the entire workflow collapsing.
- Production failure modes: Understanding why agentic systems fail in production — context overflow, instruction drift, tool hallucination, infinite loops — and designing defensively against each.
- MCP fluency: Model Context Protocol is becoming the standard for giving agents structured access to external systems. Engineers need hands-on experience building and consuming MCP servers, not just awareness that they exist.
- Claude Code proficiency: Claude Code isn't just an autocomplete tool. Engineers who've internalized it as a core development environment ship differently — they architect at a higher level and iterate faster. This shows up in the work they produce.
The gap between "knows about agentic AI" and "has shipped production agentic systems" is enormous. Most vetting processes can't detect it. Ours is built specifically to find it.
How Minimalistech Vets
Our vetting process has four checkpoints. Each one filters for something different. Together, they identify engineers who can hit the ground running on a production agentic codebase — not engineers who'll spend the first month on a learning curve.
Production Systems Shipped
We start with the portfolio question — but not the way most screeners ask it. We don't ask "have you used LangChain?" or "are you familiar with agent architectures?" We ask: show us a production agentic system you built. See our full evaluation checklist for agentic engineers for the criteria that reveal real experience versus tutorial knowledge. Walk us through the architecture. Describe what broke in the first week and how you fixed it. What's the failure mode you're most worried about? How do you monitor for agent drift? Candidates who've actually shipped answer these questions specifically. Candidates who've done tutorials go vague.
Hands-On Claude Code / MCP / LangGraph Experience
We do a live technical session — not a take-home assignment, not a whiteboard algorithm problem. We give candidates a real task: architect an agent that accomplishes a specific goal using a set of available tools. We watch how they think. Do they reach for Claude Code as a first-class development environment or as an accessory? Do they understand MCP as a protocol or just as a buzzword? Do they know when LangGraph's explicit state management is worth the overhead vs. when a simpler loop is better? The answers reveal actual experience vs. researched terminology.
Debugging Agentic Failure Modes
We show candidates broken agentic systems and ask them to diagnose the problem. This is the most revealing part of the process. Engineers who've operated production agentic systems immediately recognize patterns: this looks like context window overflow mid-conversation, this looks like a tool response that's confusing the model, this looks like a prompt that's too ambiguous for the agent to recover from. Engineers who've only studied the theory struggle to diagnose real failures. There's no substitute for having seen a system fall apart in the wild.
Architecture Judgment
The last checkpoint is a scenario-based discussion: given a specific product requirement, how would you architect this? The answer reveals judgment on the hard decisions — when to use a single agent vs. a multi-agent pipeline, how to select and scope tools so the model uses them reliably, how to manage context across long tasks without losing coherence, and when AI isn't the right solution at all. Strong candidates make confident, reasoned decisions. They explain the trade-offs. They know what they'd prototype first and why. Weak candidates hedge everything or default to the most complex possible architecture.
Why Traditional Staffing Firms Miss This
This isn't a capability problem — it's a criteria problem. Traditional staffing firms screen for "AI/ML experience," and they're doing it correctly by their standards. "AI/ML experience" means TensorFlow. It means PyTorch. It means ML engineering: training models, optimizing pipelines, deploying inference infrastructure.
That is a completely different skill set from agentic engineering.
| What traditional firms screen for | What agentic engineering requires |
|---|---|
| TensorFlow / PyTorch proficiency | Claude Code / LangGraph / agent SDK fluency |
| Model training and fine-tuning | Prompt architecture and context management |
| ML pipeline optimization | Agentic workflow reliability and failure handling |
| Data science background | Systems engineering mindset applied to AI |
| Kaggle / academic ML credentials | Production agentic systems shipped and maintained |
A senior ML engineer from a top lab might struggle to build a reliable Claude Code workflow. A mid-level backend developer who discovered agent architectures a year ago and has been obsessively building with them might deliver far more value on your product team. Traditional firms sort by the wrong axis entirely.
Beyond criteria, there's also the evaluator problem. Most staffing firms use generalist technical screeners — smart developers who can assess syntax and conventional system design but haven't personally built production agentic systems. They can tell you if a candidate knows how to write Python. They can't tell you if a candidate understands why an agent loop they designed will fail at 3am when it hits an unexpected tool response. That judgment requires lived experience with agentic systems.
At Minimalistech, every screener has shipped production agentic systems. If you want to understand the cost of getting this wrong, read our breakdown of the hidden cost of hiring the wrong AI engineer. We evaluate from the inside out.
Scaling an agentic AI team? We should talk.
Minimalistech places senior Claude Code engineers in 3–5 days. Sourced from AI-native communities. Vetted by practitioners who've built exactly what you're building.
Tell us what you need →