AI Agent Engineer (Skills & Prompts)

AI Agent Engineer (Skills & Prompts) Employment Type: Full-Time Location: Remote, Pakistan The Role Pixalate runs AI agents in production against real customer problems: agents that investigate invalid traffic (IVT) for clients, and analytics agents that let customers ask questions of our data directly.

These agents are only as good as the instructions and skills behind them.

As an AI Agent Engineer you own that layer.

You will write and maintain the system prompts and the agent skills that decide how our agents behave: what they know, which tools they reach for, when they push back, and, critically, when they stop and ask the user a clarifying question instead of guessing.

This is engineering work, not prompt tinkering.

You will version your changes, measure them against evaluations, and be accountable for agent behaviour in front of real customers.

You will work directly with the team building the IQ Agent and Analytics MCP, the client IVT investigator agents, and the MRT MCP.

Build and Own Agent Skills Author and maintain Claude skills and plugins that define agent capabilities and behaviour Design agents that ask clarifying questions and gather requirements interactively, rather than answering one-shot from incomplete information Decide what belongs in a skill, what belongs in a system prompt, and what belongs in a tool Maintain the skill library as products evolve, treating it as a living surface rather than a one-time build System Prompts for Production Agents Own the system prompts behind Pixalate's client-facing agents: IVT investigator agents, IQ Agent / Analytics MCP, and MRT MCP Version prompts and skills deliberately, with a clear record of what changed and why Debug real agent failures: wrong tool selection, hallucinated findings, missed clarifying questions, context exhaustion Evaluation and Measurement Define what "correct" means for each agent and build evaluation suites that test it Build ground-truth datasets from real customer scenarios and measure changes against them Catch regressions before customers do Context and Tool Design Engineer context strategies: what the agent sees, in what order, within what budget Work with engineering on MCP tool surfaces: which tools exist, how they are described, and how those descriptions drive correct selection Tune for cost and latency as well as accuracy Safety and Trust Guard against prompt injection, data leakage, and unsafe tool use in customer-facing agents Ensure agents represent Pixalate's findings accurately and do not overstate confidence What We're Looking For Demonstrable experience building LLM agents that shipped to real users, not prototypes or demos Hands-on work with agent skills, plugins, system prompts, or equivalent agent-behaviour tooling Evidence of evaluation discipline: you can describe an eval suite you built, the ground truth behind it, and a regression it caught. This is the single most important qualification. Strong written English. This role is writing, and the writing is the product. Working proficiency in Python or TypeScript, and comfort reading the codebase your agents run in Understanding of tool-calling / function-calling and how agents select tools Awareness of prompt injection, PII handling, and the failure modes of customer-facing AI Comfortable with some U.S. timezone overlap for team meetings Ad-tech, digital advertising, or fraud-detection exposure is strongly preferred: programmatic, DSPs, SSPs, CTV, or IVT and ad fraud concepts. If you do not have it, you must be prepared to learn the domain deeply in your first 90 days, because agent quality depends on it.