Hire agentic AI developers who have shipped long-running agents to production.

2025 was the year agentic patterns crossed from demos into real products. Claude Code, Cursor's agent mode, v0 auto-mode, Devin, and a first wave of enterprise agent deployments moved the category from "interesting notebook" to "line item on the P&L". By Q1 2026, every scale-up we talk to has at least one agentic workload in production or in a 90-day pilot, and most of them still do not have a dedicated owner for it. That is the role we fill.

An agentic AI developer is the person who owns the orchestration and control-flow layer of an LLM-backed product. They are not the LLM engineer (who owns model selection, prompting, evals) and they are not the RAG engineer (who owns retrieval and context assembly). They are the engineer who turns a single-turn completion into a multi-step, tool-using, memory-retaining, self-correcting agent that survives contact with a production user — and, critically, knows when to escalate to a human instead of spinning in a loop for 40 tool calls.

The stack in 2026 has stabilised around four primitives. Planning and orchestration frameworks (LangGraph, Microsoft AutoGen, CrewAI, and increasingly the Vercel AI SDK v6 agent patterns on the TypeScript side). First-party agent SDKs from the model labs — Anthropic's Claude Agent SDK and OpenAI's Assistants / Responses API. A tool-interface layer, which in practice means Model Context Protocol (MCP) — the spec Anthropic published in late 2024 and that has now been adopted by every major vendor. And an observability layer: LangSmith, Langfuse, or Braintrust traces, because an agent without structured traces is a production outage waiting to happen.

Anthropic's own engineering team described the design space well in their "Building effective agents" post (Dec 2024): the honest answer for most production use cases is still a workflow — a carefully composed chain of LLM calls with deterministic structure — and only a minority of problems genuinely justify an agent with open-ended tool use and a loop. A good agentic developer knows the difference, and more often talks clients out of the open-loop agent than into one. That is a hiring signal, not a concession.

Production signals we screen for are concrete and measurable: agent success rate on multi-step tasks, tool-call success rate (the share of tool invocations that return usable output on the first try), mean tokens per completed task, timeout and retry rate, and escalation-to-human rate. Candidates who cannot articulate these in concrete numbers from a system they have operated — not prototyped, operated — do not make shortlist. The SWE-bench Verified leaderboard is useful context here: state-of-the-art agent scaffolds hit ~70% in 2025, but production deployments on narrower domains routinely land at 30–50% task completion. That gap is where the real engineering work lives.

Generic "AI engineer" pipelines produce noise for this role. LinkedIn searches for "agentic AI" return thousands of candidates whose most advanced deliverable is a LangChain notebook with three tools and a single-turn call. Our pipeline is built to separate the demo-builders from the operators inside the first 15 minutes, not in the third interview round.

We source across five channels specific to agent engineering: the contributor graphs of the orchestration frameworks themselves (LangGraph, AutoGen, CrewAI, the Vercel AI SDK agents module, and the open Claude Agent SDK examples repo); MCP server authors — anyone who has shipped and maintained a non-trivial MCP server on the Anthropic or community registries; SWE-bench and GAIA leaderboard submitters, filtered to engineers (not researchers) who shipped a scaffold others could reproduce; the OpenAI Assistants / Responses API cookbook contributors; and a private CEE network of ~640 AI engineers built by Nikita during his time at Neurons Lab, where the orchestration and ML platform work overlapped heavily with today's agent stack.

Every candidate goes through a 14-minute AI technical interview that probes production agent signals. Representative questions we ask, all with adaptive follow-ups: "walk me through the last time one of your agents got stuck in a retry loop — what did the trace look like, and what was the fix?", "how do you bound an agent's tool-call budget, and how did you pick the budget?", "what is your tool-call success rate, and how did you get it there?", "when would you reach for a deterministic workflow over an open-loop agent, and give me a recent example where you made that call?". These questions are unfakeable without production experience — a candidate who has only built demos cannot invent a p95 or an escalation rate that stands up to follow-up.

The last layer is specific to this role: we require every agentic developer shortlisted to have shipped at least one agent that completed multi-step tasks for real users for >30 days, with traces we can inspect (LangSmith / Langfuse project link, a published post-mortem, a conference talk, or a redacted internal write-up). If the candidate cannot point us at a production artifact and talk through a real incident, they do not ship. Over 2025-Q4 and 2026-Q1, this filter rejected 86% of self-identified "agentic AI" inbound CVs.