Software Engineer - Agentic System Development

免费注册后解锁"你的匹配分 vs 高分画像"与个人技能缺口。

About the Role: We’re building AI into STARLIMS, a platform used across quality manufacturing, life sciences, public health, forensics, and environmental sciences. This role is focused on agentic systems: software that reasons over a task, calls tools, works through multiple steps, and hands the result to a person to review and approve. Our users work under strict accuracy, traceability, and validation requirements. The engineering challenge is making non-deterministic systems reliable, observable, and controllable enough to be trusted, tested, and shipped. You’ll work on both the platform and runtime our agents execute on and the production agents built on top of it. What You’ll Work On: Agent Platform Runtime (Core Focus) Design and build the runtime our agents execute on: planning and execution loops, tool calling, state management, durable execution, and failure recovery Build the layer through which agents reach platform data and external systems safely Design coordination, delegation, and handoff across agents and workflows where needed Make agent behavior versionable, testable, measurable, and regression-safe across releases Build reusable primitives so new agents are configured rather than rebuilt from scratch Building Agents (Core Focus) Take a domain workflow from expert conversation to a working agent: goals, actions, execution flow, failure handling, and success criteria Ground agent decisions and outputs in authoritative enterprise data rather than relying on model knowledge alone Implement human-in-the-loop by design, including approval gates, override capture, uncertainty handling, and clear evidence for agent decisions. Agents recommend and draft; people decide Close the loop: turn user corrections and overrides into signals that measurably improve the agent Evaluation Reliability Build evaluation harnesses for multi-step behavior, not single-response accuracy: task completion, tool-call correctness, groundedness, trajectory quality, and regression across model, prompt, and tool changes Define production metrics for agent quality, reliability, latency, cost, and human intervention rates Implement guardrails, fallbacks, timeouts, cost ceilings, and end-to-end observability and tracing across agent runs Design safeguards against prompt injection, unsafe tool use, excessive permissions, data leakage, and other agent-specific security risks Manage prompt evolution, model drift, and non-determinism while maintaining consistent, measurable system behavior across releases Integration Data Integrate agents with platform APIs and third-party enterprise systems already running in our customers’ environments Build retrieval and context pipelines that turn fragmented enterprise data into reliable, permission-aware agent context Design controlled execution paths for automated actions, with a complete, traceable audit trail Platform Infrastructure Build and operate backend services on AWS (Lambda, API Gateway, DynamoDB, Step Functions, etc.) Own significant parts of the system architecture and contribute to key technical decisions Contribute to infrastructure-as-code and deployment pipelines Tech Stack Languages: TypeScript, Python Backend: Node.js, Python, AWS Lambda, Step Functions AI: OpenAI, Anthropic, MCP and related agent/tool protocols, embeddings and vector search Frontend: React, Next.js, Tailwind CSS Infrastructure: AWS, Terraform Testing: Jest, Playwright, pytest What We’re Looking For Must Have 6+ years of software engineering experience, including production systems Strong understanding of LLM behavior, limitations, and failure modes, especially how errors compound across a multi-step run Experience with LLM APIs, tool and function calling, and designing planning and execution loops Experience evaluating and debugging non-deterministic systems Solid backend and cloud experience (AWS or equivalent) Proficiency in TypeScript and/or Python You Should Be Comfortable With Debugging across distributed and non-determin

关于该职位：我们正在将AI构建到STARLIMS中，这是一个广泛应用于质量制造、生命科学、公共卫生、法医学和环境科学的平台。该职位专注于代理系统：能够对任务进行推理、调用工具、执行多步骤操作，并将结果交由人工审核和批准的软件。我们的用户在严格的准确性、可追溯性和验证要求下工作。工程挑战在于使非确定性系统足够可靠、可观测和可控，从而能够被信任、测试和交付。你将同时负责代理执行所依托的平台和运行时，以及构建在其之上的生产代理。你将参与的工作：代理平台运行时（核心重点）设计和构建代理执行所依托的运行时：规划和执行循环、工具调用、状态管理、持久执行和故障恢复构建代理安全访问平台数据和外部系统的层在需要时设计代理和工作流之间的协调、委托和交接使代理行为可版本化、可测试、可度量，并在各版本间具备回归安全性构建可复用的原语，使新代理通过配置而非从头重建来创建构建代理（核心重点）将领域工作流从专家对话转化为可工作的代理：目标、操作、执行流程、故障处理和成功标准将代理决策和输出建立在权威企业数据之上，而非仅依赖模型知识通过设计实现人在回路，包括审批关卡、覆盖捕获、不确定性处理和代理决策的清晰证据。代理负责推荐和起草；由人做决定闭环：将用户纠正和覆盖转化为可度量地改进代理的信号评估可靠性构建面向多步骤行为的评估框架，而非单响应准确性：任务完成度、工具调用正确性、有据性、轨迹质量，以及跨模型、提示和工具变更的回归定义代理质量、可靠性、延迟、成本和人工干预率的生产指标实现护栏、回退、超时、成本上限，以及跨代理运行的端到端可观测性和追踪设计针对提示注入、不安全工具使用、过度权限、数据泄露和其他代理特定安全风险的防护措施管理提示演进、模型漂移和非确定性，同时在各版本间保持一致、可度量的系统行为集成数据将代理与平台API以及已在客户环境中运行的第三方企业系统集成构建检索和上下文管道，将碎片化的企业数据转化为可靠、权限感知的代理上下文为自动化操作设计受控执行路径，并具备完整、可追溯的审计轨迹平台基础设施在AWS上构建和运营后端服务（Lambda、API Gateway、DynamoDB、Step Functions等）拥有系统架构的重要部分并为关键技术决策做出贡献参与基础设施即代码和部署管道技术栈语言：TypeScript、Python后端：Node.js、Python、AWS Lambda、Step Functions AI：OpenAI、Anthropic、MCP及相关代理/工具协议、嵌入和向量搜索前端：React、Next.js、Tailwind CSS基础设施：AWS、Terraform测试：Jest、Playwright、pytest我们期望的人选必备条件6年以上软件工程经验，包括生产系统深入理解LLM的行为、局限性和故障模式，特别是错误如何在多步骤运行中累积具有LLM API、工具和函数调用经验，以及设计规划和执行循环的经验具有评估和调试非确定性系统的经验扎实的后端和云经验（AWS或同等）精通TypeScript和/或Python你应能自如应对跨分布式和非确定性系统的调试

⚑ 岗位信息收录自公开招聘页（Himalayas），以雇主官网为准