Internship - Build an agentic software refactoring workflow

免费注册后解锁"你的匹配分 vs 高分画像"与个人技能缺口。

Work Schedule
Standard (Mon-Fri)
Environmental Conditions
Office
Job Description

Problem — Why do this?

Software development requires substantial effort across requirements, design, coding, testing and review. Repeated handovers create overhead, inconsistencies and rework. AI can reduce this effort, but code generation alone is insufficient. Software must meet user needs, engineering standards and quality expectations under normal and adverse conditions. We want to test whether specialized AI agents can handle the full V-model through coordinated work, recurring feedback and human oversight.

Goal and value

Build a proof of concept that demonstrates the complete multi-agentic workflow on a concrete software improvement and measure productivity gains against quality, human effort and operating costs. A faster output using AI only matters if it meets the agreed standards. As such, the expected benefits are:

Reduce routine specification, implementation, testing and review effort.
Give engineers more time for requirements, architecture, design and quality decisions.
Build practical AI experience and stay ahead in software engineering.
Establish evidence for wider adoption.

Assignment

An intern with strong software engineering and great motivation and enthusiasm to work with Artificial Intelligence (AI) in SW development will design, implement and evaluate specialized agents for requirements analysis, functional specification, high-level design, detailed design, coding, unit testing, integration testing, system testing and user acceptance testing. Agents must improve their outputs iteratively, exchange feedback across stages and meet defined completion criteria before human-approved handover.

Scope: Refactor Acquisition Service Tool and/or Reference Image Manager to improve maintainability, reuse and extensibility. Enable feature migration and reuse across AST, RIM, Sherpa and MiCo, supporting the retirement of Acquisition Monitoring as Service Tool.
Duration: six months, supported by a software engineering mentor and a testing or quality representative.
Deliverables: A working, repeatable workflow and demonstrated software improvement: The architecture of AST has been significantly improved so that it is made more modular, easy to maintain, reusable across different projects and extendable, so that adding extra new functionality can be done faster and more efficiently by the SW teams.
The current functionality and behavior available in AST is maintained, that is, no difference with current solution.
Extensions to the existing testing suit as needed to secure zero regression impact.

Agent instructions, skills, context packages, memory and handover rules.
Linked requirements, designs, code, test plans and execution evidence, meeting the quality expectations and the internal definition of done Any gaps in requirements, expected behavior, user needs, test plans, test coverage and alike are also identified and raised to the SW teams for them to take action

Comparison with a human-driven baseline and rollout recommendations.

How — Proposed solution

Plan and coordinate. Define tasks, dependencies, agent responsibilities, outputs, completion criteria and human approval gates. An orchestrator routes work, tracks progress and manages changes. Develop test plans alongside requirements and designs.
Provide context and tools. Use existing team requirements, designs, engineering decisions, code examples, standards and known defects. Give agents controlled access to repositories, build tools, automated checks and test environments. Log actions. Shared, versioned memory retains artifacts, decisions and evidence; separate drafts from approved content.
Iterate and check. Each stage follows: read inputs → produce → check → revise → human approval → handover. Agents assess results and adapt using compilation, static analysis, tests and peer feedback. Set iteration, time and cost limits. Escalate ambiguity, repeated failures and decisions outside agent authority.
Communicate and trace. Handovers include versions, requirement references, check results, open questions and approval status. Return failures to the responsible stage and recheck affected downstream work. Changes to requirements or acceptance criteria require human approval.
Keep humans accountable. Humans define needs, use cases, behavior, design constraints, coding standards, guardrails, testing goals and quality expectations. They approve decisions and handovers, resolve ambiguity, and own verification, validation and acceptance.
Test adverse conditions. Cover normal operation, boundaries, invalid inputs, dependency failures and recovery. Base acceptance checks on approved expectations; define selected checks independently of the coding agent.

Results

Demonstrate maintainable software that meets acceptance criteria. Measure delivery time, human effort, rework, defects, requirement coverage, agent retries, interventions, costs and recovery from changes. The final recommendation must identify what to scale, what to improve and where human expertise remains essential.

工作时间
标准（周一至周五）
环境条件
办公室
职位描述

问题——为什么做这件事？
软件开发需要在需求、设计、编码、测试和评审上投入大量精力。反复的交接会带来开销、不一致和返工。AI可以降低这些投入，但仅靠代码生成是不够的。软件必须在正常和不利条件下满足用户需求、工程标准和质量期望。我们希望验证专业化AI智能体能否通过协同工作、循环反馈和人工监督来处理完整的V模型。

目标与价值
构建一个概念验证，在具体的软件改进上展示完整的多智能体工作流，并衡量生产力提升相对于质量、人工投入和运营成本的表现。使用AI更快地产出，只有在满足约定标准时才有意义。因此，预期收益是：
减少常规的规格说明、实现、测试和评审投入。
让工程师有更多时间用于需求、架构、设计和质量决策。
积累实用的AI经验，在软件工程领域保持领先。
为更广泛的采用建立证据。

任务
一名具备扎实软件工程能力、并对在软件开发中应用人工智能（AI）充满动力和热情的实习生，将设计、实现和评估用于需求分析、功能规格说明、高层设计、详细设计、编码、单元测试、集成测试、系统测试和用户验收测试的专业化智能体。智能体必须迭代改进其输出，在各阶段之间交换反馈，并在人工批准的交接之前满足定义的完成标准。

范围：重构Acquisition Service Tool和/或Reference Image Manager，以提升可维护性、复用性和可扩展性。实现功能在AST、RIM、Sherpa和MiCo之间的迁移与复用，支持Acquisition Monitoring as Service Tool的退役。
持续时间：六个月，由一名软件工程导师和一名测试或质量代表提供支持。
交付物：一个可运行、可重复的工作流，以及经过演示的软件改进：AST的架构已得到显著改进，使其更加模块化、易于维护、可跨不同项目复用且可扩展，从而软件团队能够更快、更高效地添加额外的新功能。
AST中当前可用的功能和行为得到保持，即与当前解决方案无差异。
根据需要扩展现有测试套件，以确保零回归影响。

智能体指令、技能、上下文包、记忆和交接规则。
关联的需求、设计、代码、测试计划和执行证据，满足质量期望和内部完成定义。需求、预期行为、用户需求、测试计划、测试覆盖率等方面的任何缺口也会被识别并上报给软件团队，由其采取行动。

与人工驱动基线的对比及推广建议。

如何做——建议的解决方案
计划与协调。定义任务、依赖关系、智能体职责、输出、完成标准和人工批准关口。由编排器路由工作、跟踪进度并管理变更。测试计划与需求和设计同步制定。
提供上下文和工具。使用团队现有的需求、设计、工程决策、代码示例、标准和已知缺陷。为智能体提供对代码仓库、构建工具、自动化检查和测试环境的受控访问。记录操作日志。共享的、带版本控制的记忆保留工件、决策和证据；将草稿与已批准内容分开。
迭代与检查。每个阶段遵循：读取输入→产出→检查→修订→人工批准→交接。智能体通过编译、静态分析、测试和同行反馈来评估结果并进行调整。设定迭代、时间和成本限制。对歧义、反复失败和超出智能体权限的决策进行上报。
沟通与追溯。交接内容包括版本、需求引用、检查结果、未决问题和批准状态。将失败返回给责任阶段，并重新检查受影响的下游工作。对需求或验收标准的变更需要人工批准。
保持人类问责。人类定义需求、用例、行为、设计约束、编码标准、护栏、测试目标和质量期望。他们批准决策和交接、解决歧义，并负责验证、确认和验收。
测试不利条件。覆盖正常运行、边界、无效输入、依赖失败和恢复。验收检查基于已批准的期望；部分检查独立于编码智能体定义。

结果
展示满足验收标准的可维护软件。衡量交付时间、人工投入、返工、缺陷、需求覆盖率、智能体重试次数、干预次数、成本和从变更中恢复的能力。最终建议必须指出哪些方面应扩展、哪些方面应改进，以及哪些方面仍然需要人类专业知识。

⚑ 岗位信息收录自公开招聘页（ATS 官方），以雇主官网为准