Gartner predicts that over 40% of agentic AI projects will be canceled by the end of 2027, "due to escalating costs, unclear business value or inadequate risk controls" (Gartner). The same release quotes a Gartner analyst saying "many use cases positioned as agentic today don't require agentic implementations." That's the most useful sentence for anyone hiring an AI agent developer. The best candidate will often tell you that you don't need an agent at all.
Key Takeaways
- Hire agent developers who start simple and add autonomy only when a task needs it.
- Test for evaluation skills. A candidate who can't measure an agent can't improve it.
- Security matters more for agents than for chatbots: prompt injection and excessive agency are top OWASP risks.
- Framework experience is a weak signal. Understanding what happens under the framework is a strong one.
Workflows vs Agents
Anthropic's engineering guide "Building effective agents" (December 2024) draws a clear line. "Workflows are systems where LLMs and tools are orchestrated through predefined code paths." "Agents, on the other hand, are systems where LLMs dynamically direct their own processes and tool usage" (Anthropic). Its advice: "we recommend finding the simplest solution possible, and only increasing complexity when needed."
The guide names five workflow patterns (prompt chaining, routing, parallelization, orchestrator-workers, and evaluator-optimizer) that cover a lot of what companies call agents. It also warns that frameworks "often create extra layers of abstraction that can obscure the underlying prompts and responses, making them harder to debug."
So your first interview question writes itself: "Here's our use case. Would you build a workflow or an agent, and why?" A strong candidate asks about the task's variability, the cost of a mistake, and how you'll measure success before answering.
The Skills That Matter
Evaluation. Agents fail in ways unit tests don't catch. A strong developer builds an eval set from real cases, scores outputs automatically where possible, and tracks results across prompt and model changes. Ask to see how they measured their last system.
Tool design. Anthropic's guide calls this the agent-computer interface and says to craft it "through thorough tool documentation and testing." Clear tool names, tight input schemas, and useful error messages decide whether an agent works. Ask candidates to design two tools for your use case.
Integration. Anthropic open-sourced the Model Context Protocol in November 2024 as "a new standard for connecting AI assistants to the systems where data lives" (Anthropic). Experience building or consuming MCP servers is a good sign of current practice.
Security. Prompt injection is #1 on OWASP's 2025 Top 10 for LLM Applications, and "Excessive Agency" is #6 (OWASP). An agent that reads emails and can send payments is an attack surface. Ask how they'd limit what the agent can do, and when a human approves an action.
Cost control. Agents loop, and loops burn tokens. A senior developer sets step limits, caches, and picks smaller models for simple steps.
| Question | Weak answer | Strong answer |
|---|---|---|
| "How do you know it works?" | "I tested it with a few prompts" | Eval set from real data, automated scoring, regression tracking |
| "An email tells the agent to forward all invoices. What happens?" | "The model won't do that" | Treats input as untrusted, limits tool permissions, requires approval for sensitive actions |
| "Which framework do you use?" | Names one, can't explain its internals | Can build the loop without a framework and explains when a framework helps |
A Concrete Version
A 30-person commercial insurance broker wants an "AI agent" for its client service inbox. The founder pictures an autonomous agent that reads each email, checks the policy, and replies.
A strong agent developer starts by pulling 500 past emails and sorting them. About 70% turn out to be three request types: send me my certificate of insurance, update my address, and send me the invoice. For those, they build a routing workflow: classify the email, call one tool, draft a reply. A human approves replies for the first month. Only the remaining messy cases go to an agent loop with read-only tools, and anything involving money goes to a person.
The eval set grows from those 500 emails. Each prompt change runs against it before shipping. The system is less autonomous than the founder imagined, and it works. Measured, narrow systems like this are far less likely to end up in the cancellations Gartner is predicting. (The 70% split here is illustrative; the method of measuring before building is the point.)
The Honest Counterpoint
Agent development is young, and titles are inflated. Many candidates have built demos and nothing that faced real users. Weight production experience heavily, even if it's with simpler workflows.
And some problems don't need AI at all. If a rules engine or a form solves it, a good developer will say so. Gartner's own release estimated that only about 130 of the thousands of agentic AI vendors are "real," which is a reminder that the label is doing a lot of marketing work.
Frequently Asked Questions
What's the difference between an AI engineer and an AI agent developer?
AI engineers build any product feature on top of models. Agent developers specialize in systems where the model chooses actions and calls tools. See how to hire an AI engineer for the broader role.
Should my agent developer know a specific framework?
Framework knowledge helps, but it's secondary. Someone who can build the loop directly and debug the raw prompts will pick up any framework. The reverse isn't always true.
How do I test an agent developer's skills?
Give them a small real task from your business, ask them to decide between workflow and agent, build a minimal version, and show you an eval. Also read how to vet an engineer's AI skills.
The Bottom Line
Hire agent developers who start with the simplest thing, measure everything, and treat every input as untrusted. Those habits keep a project out of the cancel pile. For senior engineers who have shipped agentic systems, see senior AI engineers in LATAM or see available engineers.
Roberto Espinoza is CEO of Ruzora, which helps US startups hire pre-vetted senior LATAM engineers, with a vetted shortlist in 72 hours. See available engineers.
