AI/Agent Engineer — Model Gateway & Agent Runtime

About the Role:
You’ll build the core of how our platform talks to and constrains large language models. This isn’t prompt engineering — it’s building the actual system boundary that keeps an AI agent from ever executing a real action directly. You own the model gateway (the single door through which all agent logic reaches the model) and the agent runtime (the loop that decides what a model’s proposal is actually allowed to do).

Responsibilities:

• Build the model gateway: the single module through which all agent logic reaches the LLM provider, with structured-output validation at the boundary.
• Design the gateway so switching model providers or endpoints is a configuration change, not a code change.
• Build the agent runtime: the propose → validate → execute-or-hold → record loop that enforces what a model is permitted to do.
• Implement tool-call validation against an allowlist and risk-category policy, before any real action executes.
• Implement hard-stop budget enforcement (steps, tokens, cost, wall-time) and repeated-action-loop detection.

Requirements:

• 3+ years of production LLM integration experience.
• Strong experience with structured-output / tool-use patterns (schema-validated model outputs).
• A systems-safety mindset — you think about failure modes and boundaries, not just prompt quality.

Nice to Have:

• Experience building agentic systems where the model proposes actions but code (not the model) makes the final call.
• Experience with the Anthropic API or comparable LLM provider APIs in production.

Work Location: In person