Agent Harness Engineer

About Axiom:

Axiom is building the closed-loop scientific AI system required to replace animal testing and, over time, much of human safety testing. We start with pharma’s hardest drug development toxicology problems. Those problems define the proprietary human biological data we generate through Axiom’s Data Factory. We use that data to train scientific AI, partnering with leading AI labs to improve frontier models while building our own specialist agentic harness to deploy the improved frontier models back into pharma. Each deployment reveals the next capabilities to build, creating a compounding loop across data, models, and drug development. Today, liver toxicity is our proving ground. Axiom is already helping leading pharmaceutical companies understand toxicity, identify its mechanism, and design safer drugs. Over time, we will expand across the major organ systems and build the experimental and agentic system of record for translational drug development. Our goal is to dramatically reduce the risk of testing new molecules in humans, enabling high throughput evaluation of efficacy in humans.

What you will do:

• Own the harness: the scaffolding, tooling, and infrastructure that turn frontier models into agents that do long-horizon scientific analysis

• Build the data backbone that gets everything to the right place: pipelines, storage, and systems for runtime context, agent trajectories, eval results, and training data

• Build sandboxed execution environments with instant spin-up/tear-down, reproducible and deterministic enough to trust for evals and RL

• Design and run the eval systems: offline suites, test cases on production traces, LLM-as-judge pipelines, regression gates

• Work closely with domain experts and encode their taste into rubrics, golden sets, and review workflows, turning "I know it when I see it" into something measurable

• Build the tools the agent needs to do better work, and stay tuned in to the state of the art for new methods, protocols, and patterns worth adopting

• Make every agent run observable and replayable: trace every model call, tool call, and state transition, and build the debugging tooling to make sense of it

• Engineer the context: memory, compaction, retrieval, and recovery so long-horizon agent runs stay coherent across hours and crashes

• Own the loop itself: retries, budget caps, stop conditions, output verification, permissions, and guardrails

• Support ML research with environments, reward instrumentation, and rollout infra for RL on agentic tasks

Expertise:

• Python, Modal, DuckDB, FastAPI, Docker, Containerization, Terraform

• Engineers who've built with LLM APIs and shipped agentic systems: tool use, loops, and the debugging scars to prove it

• Built bespoke evaluation, monitoring, and RL env observability tooling (SvelteKit, Svelte 5, React)

What we look for:

• Can tackle deep technical challenges and own/ship simple, clean, maintainable code

• High ownership: owns outcomes end to end, not tickets, and doesn't wait for a spec to start moving

• Allergic to complexity: reaches for the simplest system that works and keeps it that way as it scales

• Strong software engineer first, with infrastructure, platform, data, or devtools depth and production systems they're proud of

• Instinctively asks "how would we know if this is working?" and builds the measurement alongside the feature

• Reads an agent failure trace the way other engineers read a stack trace

• Cares about reliability because they know environments that break silently poison evals and training data

• Has a knack for surfacing the important questions about what the agent actually needs to do better work

• Sharp and confident

  • keeps up with a really technical & sophisticated crowd

  • comfortable getting in over their head and figuring it out as they go

  • thrives in a discipline with no playbook, because it's being invented right now

• Curious about how things work: an engineering/tinkering mindset, good at scavenging the state of the art

• Passion for learning what "good" looks like from deep domain experts and turning it into systems