Agent Engineer II

Amigo•New York, NY

25d

About The Position

As an Agent Engineer II at Amigo, you'll independently design and implement production AI agents for healthcare customers. You'll architect context graphs that model complex clinical workflows, design agent personalities that maintain clinical safety, and build evaluation frameworks that catch problems before patients encounter them. This role requires you to make design tradeoffs -- balancing conversation quality, clinical safety, and system reliability -- with minimal oversight.

Requirements

2-4 years of production software engineering experience
Strong Python skills including Pydantic models, async patterns, and building reliable systems that interact with external APIs
Experience with LLMs, prompt engineering, or building on AI platforms
Ability to design systems by reasoning about competing constraints -- you understand that boundary constraints matter more than action guidelines, and that quality trumps speed
Experience working directly with customers or domain experts to translate requirements into technical implementations
Debugging skills across multiple system layers -- you can trace a problem from user-visible symptom to root cause across logs, prompts, and configuration
Understanding of testing methodologies -- you think about what to measure, not just whether tests pass
Clear technical communication for both engineering and clinical audiences

Nice To Haves

Experience in regulated industries (healthcare, finance, legal)
Background with state machine design, finite automata, or conversation flow modeling
Experience with simulation frameworks or synthetic data generation
Understanding of distributed systems and observability (Datadog, structured logging)
Familiarity with compliance requirements (HIPAA, SOC 2)
Experience with voice/TTS systems and audio-specific constraints

Responsibilities

Design context graphs (hierarchical state machines) that model multi-step clinical workflows -- choosing between linear arcs and routing hubs, calibrating state density, and preventing conversation loops
Architect agent identities: background, motivations, expertise, behaviors, and communication patterns that produce clinically safe and engaging conversations
Build dynamic behavior sets that inject contextual instructions at runtime -- designing trigger conditions, choosing override modes, and testing activation patterns
Design user memory systems by defining extraction dimensions that are bounded, orthogonal, and actionable -- preventing the dimension overlap and storage explosion that collapse memory systems
Write tool integration specs that define when tools fire, what parameters they receive, and how results persist in conversation context
Diagnose production conversation failures by reading prompt logs, tracing routing decisions, and identifying root causes across the agent-graph-behavior stack
Design evaluation suites: metrics that resist gaming (Goodhart's Law), personas that represent real patient populations, and scenarios that test edge cases
Run coverage-optimized simulations using frontier and heatmap algorithms to systematically test all reachable states and transitions
Process complex customer feedback -- categorizing issues into agent design problems, context graph flow issues, platform bugs, and knowledge gaps