A lot of AI agent developer job descriptions ask for something that doesn't exist. The usual example: five or more years of experience building AI agents.
Next to other senior engineering roles, that number looks reasonable. It isn't. The technology behind today's agents is a few years old, so a requirement like that backfires. It scares off the honest candidates you want and rewards the ones willing to stretch the truth.
Here's how much agent experience is realistic, what inflated requirements cost you, and how to write a description that rewards the engineering depth that predicts success.
How much agent experience can anyone have?
Most agents run on an idea from the ReAct paper, published in October 2022: a loop where a model reasons, takes an action, and looks at the result. The commercial pieces came later. OpenAI added function calling to its API in June 2023, and Google brought it to Gemini that December. Anthropic made tool use generally available in May 2024, and the Model Context Protocol, now the standard way agents connect to tools, followed in November 2024.
So as of late 2026, nobody has more than about three and a half years of hands-on experience with commercial tool calling. Most production agents went live in 2024 or later. A strong senior candidate today usually has two to three years of real agent and LLM production work, sitting on top of many years of general software engineering.
For a point of comparison, LinkedIn's 2026 Jobs on the Rise report named AI engineer the fastest-growing job title in the U.S. The median experience for that broader role was 3.7 years.
Why inflated requirements backfire
Ask for the impossible and you change who applies.
Experienced engineers see "5+ years of AI agent experience" and decide the hiring team doesn't understand the field, or that the role isn't for them. Many of your best candidates, the backend and platform engineers who've spent two years putting agents into production without fanfare, won't apply at all.
The people who do apply are the ones happy to count every machine learning project, chatbot, or automation script as "agent experience." Gartner calls the vendor version of this "agent washing," where assistants, chatbots, and robotic process automation get rebranded as agentic AI without much agentic capability behind them. Resumes get washed the same way.
The search slows down too. Recruiters screen out good people based on years, interviewers burn time on candidates who only looked strong on paper, and the role sits open longer than it should.
What to require instead
A better job description splits experience in two. General engineering depth can be measured in years, so require it that way. Agent-specific experience should be described as evidence the candidate can show.
For a senior role, a realistic requirement reads like this: five or more years of professional software engineering, including production systems that integrate with external APIs, plus hands-on experience building and running at least one LLM-based agent or tool-using system that real users relied on.
Mid-level: three to six years of software engineering and at least one agent or LLM feature shipped to production. Junior: solid engineering fundamentals and a portfolio with original agent work. Both are fair bars.
Then spell out the skills in plain language. A strong candidate can design tools and APIs a model uses correctly and decide what the model sees at each step. They build eval sets and use them to judge whether a change is safe to release. They can trace and debug multi-step behavior in production, keep cost and latency in check, and design approval points and defenses against prompt injection.
Notice there's no framework on that list. That's on purpose, because the tools in this space turn over every year. Microsoft put AutoGen and Semantic Kernel into maintenance mode in favor of its new Agent Framework, and OpenAI shut down its Assistants API in August 2026 in favor of the Responses API. An engineer who learned the patterns on one framework can usually pick up another in a few weeks.
Weight transferable engineering depth
Some of the best agent developers come from neighboring disciplines. Your job description can invite them in instead of screening them out.
Backend and distributed systems engineers know API design, retries, idempotency, and state management, the foundations of reliable tools. Evaluation work comes naturally to QA and test engineers. From SRE you get tracing, incident response, and cost discipline. Search and data engineers understand retrieval and data quality, and ML engineers know how to measure model behavior with rigor.
You can say so outright with a line like: "We welcome engineers from backend, search, data, QA, or SRE backgrounds who have started building with LLMs and want to take agents to production." One sentence widens the pool without lowering the bar.
Describe the work, not just the stack
Often the most useful part of a job description is the problem itself. Tell candidates what the agent will do, which systems it will touch, how much autonomy it gets, and where the project stands today. A strong engineer cares more about "move our claims-intake agent from pilot to production with a measurable quality bar" than about a list of fifteen technologies.
Being clear about the project's stage also helps you pick the right engagement model. Still proving the business case? A contract engineer who has taken agents to production before can move faster and lower your risk. If the program is established and needs a long-term owner, a direct hire makes more sense.
A sample requirements section
Here's how that might look in a real senior posting. Swap in your own systems and first project.
What you'll do: Take our first customer-service agent from a working pilot to a production system that staff and customers rely on. You will own its tools and integrations with our CRM and ticketing systems, build the evaluation set we use to decide whether a change is safe to release, add tracing and cost controls, and design the points where a person approves the agent's actions.
What you'll bring: Seven or more years of professional software engineering, including production services that integrate with external APIs. Hands-on experience building and operating at least one LLM-based agent or tool-using system that real users depended on. Comfort with Python or TypeScript, asynchronous programming, and failure handling such as retries, timeouts, and idempotent operations. A habit of measuring quality with test sets rather than impressions.
Helpful but not required: Experience with an orchestration framework such as LangGraph or a provider agent SDK, building or using MCP servers, retrieval and vector search, or working with regulated data.
Look at what's left out. The posting doesn't require a framework or a degree, and it doesn't inflate agent years. Every line describes something a candidate can back up with evidence in an interview.
A short checklist before you post
Before you publish, read the description against three questions. Does any requirement ask for more agent experience than the technology has existed? Does it name a required framework where a pattern would do? And is the work described clearly enough that a strong engineer from a neighboring field could see themselves in it?
If you're unsure on any of them, revise. The right candidates read closely.
Our complete guide on how to hire AI agent developers covers skills, realistic experience levels by seniority, interview questions, and compensation planning ranges. Once applications come in, What to Look for Beyond the Resume shows how to screen GitHub profiles, portfolios, and nontraditional backgrounds, and Real vs. Resume has the interview questions that separate production experience from demo experience. The free AI Agent Developer Interview Scorecard will help you score them. For other specialized searches, browse our hard-to-hire IT roles hiring guides.
|
Want a second set of eyes on a draft? Overture Partners will review your AI agent job description and flag the requirements that are shrinking your pool. It's also the first step of our PRECISE Talent Blueprint: define the role before sourcing starts. Request a job description review. |
Sources
ReAct paper, October 6, 2022; OpenAI function calling, June 13, 2023; Gemini function calling, December 2023; Anthropic tool use general availability, May 30, 2024; Model Context Protocol, November 2024.
LinkedIn, Jobs on the Rise 2026.
Gartner press release, June 25, 2025, on "agent washing."
Microsoft Agent Framework 1.0 general availability and AutoGen/Semantic Kernel maintenance mode.
OpenAI Assistants API deprecation notice, shutdown August 26, 2026.