AI Harness Engineer

Chemstack AI needs a foundational agent harness that can reliably orchestrate models, scientific tools, state, and long-running execution across complex chemistry workflows. This hire solves the engineering bottleneck between promising agent behavior and a production system that can be evaluated, recovered, observed, and used by customers and research teams.

You will build and own the agentic harness at the core of Chemstack AI, including the runtime, execution graph, tool contracts, state management, evaluation infrastructure, and production reliability layer. The primary spike is designing and shipping a composable, inspectable, and recoverable execution system that lets models plan synthesis routes, call scientific tools, process laboratory feedback, and support rollout and trajectory collection for future post-trained models.

What You'll Own

• Design and build the agent runtime for planning, tool calls, retries, checkpoints, state, and handoffs to human experts across long-running chemistry workflows.
• Represent workflows as composable, inspectable, and recoverable graphs of agents, tools, and decisions.
• Wrap retrosynthesis engines, reaction databases, quantum chemistry and simulation packages, and laboratory systems behind clear, well-tested contracts.
• Build evaluation harnesses, tracing, and regression suites for agent behavior, measuring outcomes at the level of the chemist's task.
• Partner with AI researchers on rollout infrastructure, trajectory collection, and reward plumbing for reinforcement learning and post-training.
• Own production latency, cost, concurrency, caching, observability, and reliability.
• Work directly with chemists and product engineers to turn laboratory workflows into agent capabilities.
• Contribute to engineering standards, technical direction, mentoring, and open-source work where appropriate.

Technical Requirements

• Strong production Python engineering skills.
• Experience building an agent harness, orchestration runtime, or graph-based agent framework.
• Deep hands-on experience with tool use, function calling, and multi-agent coordination beyond prompt templates and retrieval pipelines.
• Systems expertise across state, concurrency, failure modes, idempotency, and long-running jobs.
• Experience working across model APIs, tool interfaces, and underlying infrastructure.
• Ability to design evaluations before shipping and use traces to identify agent failure points.

Execution and Ownership

• Own the harness from architecture through production.
• Build systems that are composable, inspectable, recoverable, and reliable when workflow steps fail.
• Partner closely with AI researchers, chemists, and product engineers.
• Measure agent behavior through evaluations, tracing, and regression testing.
• Shape engineering standards and technical direction as an early member of the founding AI team.

Background

• 2–6 years of relevant professional experience.
• Relevant experience may come from a frontier AI lab, an agentic AI company, or an enterprise AI company.
• Strong engineers with a master's degree or an unfinished PhD are encouraged to apply; a PhD is welcome but not required.
• A chemistry background is not required, but the role requires the ability and interest to learn the scientific domain from subject-matter experts.

Nice to Have

• Experience with RL environments, rollout systems, or trajectory data pipelines supporting large-model post-training.
• Experience building agent evaluation frameworks, tracing, or observability for LLM systems.
• Experience integrating scientific or computational tools, structured data, or knowledge graphs into agent workflows.
• Publications, open-source contributions, or deployed systems demonstrating original technical work and execution.

Who Will Thrive Here Engineers who enjoy building foundational systems rather than application-layer integrations, think rigorously about failure modes and evaluation, and are motivated to learn a new scientific domain alongside chemists and researchers will thrive in this role.