AI Agent Engineer Intern

About Lexsi Labs

Lexsi Labs is a frontier AI lab focused on building aligned, interpretable, and safe AI systems for real-world deployment. Our work spans agentic AI, alignment, interpretability, evaluation systems, enterprise AI, and foundational model research across structured, tabular, and enterprise data. We build AI systems where transparency, auditability, robustness, and controllability are first-class constraints. Our environment is fast-moving, research-driven, and highly hands-on. We value people who bring open ideas, take ownership, and turn ambiguity into meaningful outcomes.

The Role

We are looking for AI Agent Interns to work on high-impact agentic AI problems across applied, research, and product directions. This is not a narrow internship focused on building small internal bots. You will work on real agentic systems, evaluation infrastructure, enterprise use cases, research prototypes, and product tooling that can shape how AI agents are built, tested, aligned, and deployed. Depending on your interests and strengths, you may contribute across one or more of three tracks:

• Applied Agentic AI

• Work on high-impact enterprise agentic use cases.

• Build agents that reason over complex workflows, interact with tools and data systems, execute multi-step processes, evaluate outcomes, and operate reliably in real-world environments.

• Example work includes enterprise workflow agents, agents over structured or proprietary data, human-in-the-loop agent systems, decision-support agents, and applied prototypes that can evolve into production-grade components.

• Success in this track means building useful, reliable, and high-value agentic systems that solve real enterprise problems.

• Agent R&D and Evaluation

• Work on new ideas in agent architecture, harness design, evaluations, long-horizon reasoning, alignment, and interpretability.

• Build new agent harnesses, design evaluation frameworks, study agent failure modes, benchmark tool use and planning, or contribute to research papers, technical reports, and reusable research infrastructure.

• Success in this track means producing research-quality insights, evaluation systems, benchmarks, or technical contributions that improve how agentic systems are understood and assessed.

• Product and Tooling Agents

• Help build agents that improve how researchers, engineers, and users interact with Lexsi systems.

• Examples include an AlignTune Agent for alignment workflows, AI Scientist-style agents for hypothesis generation and experiment execution, tools for evaluating and debugging agents, or product prototypes that turn research capabilities into usable workflows.

• Success in this track means building tools that compound the productivity of researchers, engineers, or enterprise users.

Key Responsibilities

• Explore open-ended agentic AI problems and convert promising ideas into working prototypes, experiments, research artifacts, or product capabilities.
• Build applied agents, design evaluations, create harnesses, experiment with planning and memory architectures, analyze failures, and document findings.
• Collaborate with research, engineering, and product teams.
• The expected outcome may be a production-grade applied component, a research contribution, a paper, a benchmark, an internal tool, or a product prototype.

Requirements & Tech Stack

• Strong programming skills, preferably in Python.
• Hands-on experience with LLMs or AI systems.
• A strong interest in agents, evaluation, alignment, interpretability, or enterprise AI.
• Comfort working on ambiguous problems, building prototypes quickly, reading papers or technical material, documenting findings clearly, and independently identifying high-value directions.

Bonus Signals

• Experience with agent frameworks.
• Experience with long-horizon agents.
• Experience with evaluation pipelines.
• Experience with research projects.
• Experience with open-source tools.
• Experience with backend systems.
• Experience with enterprise data.
• Experience with AI safety and interpretability.