1. Working with Agentic AI: Machine Learning Engineer

How human skills and agentic skills can be dynamically integrated and coordinated across machine learning engineering responsibilities

1.1. Machine Learning Engineer as an Occupation

A machine learning engineer is a professional who designs, develops, evaluates, deploys, maintains, and improves machine learning systems that use data and computational models to produce predictions, classifications, recommendations, generated outputs, or other results. The occupation connects machine learning methods with software engineering, data systems, computing infrastructure, and production environments.

Machine learning engineering extends beyond developing or training models. A machine learning engineer must understand how data, model architecture, training procedures, evaluation methods, software components, deployment environments, production inputs, and operational constraints affect one another throughout the machine learning lifecycle.

Typical responsibilities of a machine learning engineer include:

• machine learning problem formulation and solution design;
• data preparation and feature engineering;
• model development, training, and experimentation;
• model evaluation, validation, and performance optimization;
• model deployment, serving, and system integration;
• machine learning pipelines, automation, and lifecycle engineering;
• production monitoring, troubleshooting, and continuous model improvement; and
• machine learning software engineering, technical coordination, and knowledge development.

These responsibilities are fulfilled through different tasks and subtasks, and each task may require multiple actions with different skill requirements. Human skills and agentic skills can both contribute to performing these responsibilities, but their levels and forms of involvement may differ according to the characteristics of the responsibility, task, action, data, technical environment, and expected outcome.

1.2. Human Skills and Agentic Skills in Machine Learning Engineering

Machine learning engineering increasingly involves both human skills and agentic skills.

Human skills are reusable capabilities possessed and developed by machine learning engineers. They may include machine learning knowledge, statistical reasoning, software engineering, data reasoning, experimental design, system understanding, problem formulation, communication, contextual interpretation, creativity, and professional judgment.

Agentic skills are reusable capabilities available to AI agents through their underlying models, contextual information, instructions, constraints, tools, and computational resources. In machine learning engineering, these skills may include information retrieval, dataset inspection, extraction, classification, summarization, comparison, calculation, code generation, test generation, experiment execution, metric analysis, documentation support, and tool use.

Both skill types can contribute to the same occupational responsibilities, but their involvement levels vary. Problem formulation may depend heavily on human understanding of objectives and context while using agentic skills for information retrieval and technical comparison. Model experimentation may involve high levels of both human and agentic skills, with the engineer forming hypotheses and the AI agent implementing and executing repeated experiments. Production monitoring may rely heavily on agentic analysis of large volumes of system information while requiring human judgment to determine the significance of detected changes.

The useful relationship is therefore not a fixed division in which certain responsibilities permanently belong to humans and others belong to AI. Human and agentic involvement can change from one task to another and even between different actions within the same task.

1.3. Dynamic Human–Agentic Skill Integration and Collaboration in Machine Learning Engineering

Machine learning engineering can be understood as an occupation in which responsibilities are fulfilled through tasks and subtasks, tasks and subtasks are executed through actions, and actions require appropriate configurations of human and agentic skills.

An occupation provides the overall professional context. For machine learning engineering, this context includes the development and operation of machine learning systems and the associated professional practices, technical expectations, and responsibilities.

A responsibility represents an expected area of professional work or an outcome that the machine learning engineer is expected to accomplish or maintain. Responsibilities such as model development, deployment, and production monitoring may each contain multiple tasks.

A task is a context-dependent unit of work used to fulfill part or all of a responsibility. When a task is too broad or complex for reliable execution, it may be decomposed into smaller subtasks.

Tasks and subtasks are executed through one or more actions. Actions represent reusable and configurable operations such as retrieving technical information, inspecting a dataset, generating code, executing an experiment, comparing model results, diagnosing a production change, or reviewing an engineering decision.

Each action has skill requirements. These requirements describe the capabilities needed for the action to be executed effectively and reliably. Depending on the action and its context, the requirements may be satisfied by human skills, agentic skills, or an integration of both.

The skills available for machine learning engineering reside in a skill space containing both human skills and agentic skills. Some skills in this space become candidate skills for executing the actions required by a particular task. A human skill in statistical interpretation, for example, may be a candidate for evaluating model behavior, while an agentic skill in metric analysis may be another candidate for related analytical actions.

The candidate skills are considered relative to the requirements of the action and the occupational and task context. Appropriate skills are then selected and organized into a skill configuration for execution. A configuration may rely mainly on human skills, mainly on agentic skills, or on an integrated arrangement of both.

1.3.1. Skills Competition and Skills Collaboration

Two important relationships may occur between human skills and agentic skills when a skill configuration is constructed: skills competition and skills collaboration.

Skills competition occurs when both a human skill and an agentic skill can satisfy the same or similar requirements for an action, but both do not need to be applied simultaneously. Under a particular occupational and task context, one may be better suited than the other.

For example, both a machine learning engineer and an AI agent may be capable of generating a standard data-validation script. When the requirements are clearly defined and the output can be easily verified, an agentic skill may provide faster and more consistent execution. When the validation logic depends on ambiguous business rules, unusual data semantics, or incomplete requirements, the engineer’s human skills may be more suitable.

The purpose of skills competition is therefore not to establish whether human or agentic skills are generally superior. It is to identify which available skill provides the better fit for a particular action under particular conditions.

Skills collaboration occurs when the requirements of a task or its actions cannot be satisfied effectively by one skill type alone, or when combining complementary human and agentic skills is expected to produce better task performance.

Collaboration does not require human and agentic skills to operate at exactly the same time. They may be activated at different stages of action execution in a sequential, iterative, or recursive manner.

For example, an engineer may define an experimental hypothesis, an AI agent may generate and execute the required experiments, the engineer may interpret an unexpected result, the agent may perform additional analysis requested by the engineer, and the engineer may then revise the original hypothesis. Human and agentic skills are repeatedly activated according to what the evolving task requires.

This dynamic relationship allows machine learning engineers to exploit the complementary strengths of both skill types rather than relying on a permanent allocation of work.

1.3.2. Skill Configuration and Task Execution

A skill configuration specifies which human skills and agentic skills will be used to execute the actions associated with a task or subtask and how those skills will be coordinated.

The configuration depends on factors such as the action requirements, available data, task ambiguity, technical complexity, expected reliability, available tools, computational resources, production consequences, and intermediate results.

For one action, an agentic skill may independently satisfy the requirements. For another, a human skill may be selected instead. A more complicated task may require several configurations across multiple actions, with human and agentic skills repeatedly interacting as the task progresses.

The configuration is therefore dynamic rather than permanent. It may change when the action changes, when new evidence becomes available, when an agentic result is unreliable, or when human interpretation identifies a new direction for the work.

Throughout this process, the machine learning engineer remains the professional responsible for promoting and controlling the work associated with the occupation and its responsibilities. Agentic AI provides powerful complementary capabilities that can expand what the professional can inspect, generate, analyze, compare, test, and execute.

1.3.3. Task Performance Evaluation and Skill Space Adaptation

The performance produced through a selected skill configuration must be evaluated against predefined evaluation metrics. These metrics determine whether the human skills, agentic skills, or their integrated configuration are sufficiently capable of performing the relevant task and its required actions.

The evaluation criteria depend on the work being performed. In machine learning engineering, they may include correctness, predictive performance, robustness, reproducibility, reliability, execution time, computational cost, latency, compatibility, maintainability, or other task-specific expectations.

When a selected skill or skill configuration produces performance above the defined threshold, it passes the performance evaluation and can be retained for future applications under appropriate conditions. Repeated successful use provides additional evidence about where that skill or configuration can be applied effectively.

When performance falls below the required threshold, the corresponding skill or configuration should not simply be reused without modification. The relevant human or agentic skills may be refined or reconfigured. Existing skills may also be combined differently, or new human skills and agentic skills may be developed or constructed from scratch when the existing skill space does not provide an adequate capability.

For example, if an AI agent repeatedly generates training pipelines that execute correctly but fail reproducibility requirements, its agentic skill may be refined through better instructions, validation procedures, tool support, or additional testing. If the engineer lacks knowledge required for a new deployment environment, the corresponding human skill may be developed through learning and practical experience.

Refined and newly constructed skills that satisfy the required performance criteria can be added to the skill space. In this way, task performance evaluation provides feedback not only for completing the current work but also for continuously optimizing and expanding the available skill space.

The skill space should not be constructed as an isolated collection of occupation-specific representations. Some human and agentic skills used in machine learning engineering are reusable across occupations. Information retrieval, statistical reasoning, technical comparison, code generation, summarization, communication, calculation, and tool use may contribute to work in software engineering, data science, scientific research, finance, healthcare, education, and many other occupations.

Reusable skills can therefore be represented once and configured differently according to the responsibilities, tasks, actions, and contexts of different occupations. This allows the skill space to serve a growing range of occupations without unnecessarily duplicating or exploding the number of skill representations. Occupational specialization can be introduced through task context, action requirements, parameters, tools, constraints, and skill configurations rather than by creating a completely separate skill representation whenever the same underlying capability appears in another occupation.

The resulting process is adaptive:

occupational responsibility → task and subtask → actions → skill requirements → candidate human and agentic skills → skill configuration → task execution → task performance evaluation → skill refinement, reconfiguration, or construction → optimized and expanded skill space

This cycle allows human and agentic capabilities to evolve together with the requirements of machine learning engineering work.

1.4. Essential Responsibilities of Machine Learning Engineer

1.4.1. Machine Learning Problem Formulation and Solution Design

Machine learning problem formulation and solution design establishes what problem should be addressed, what outcome the system should produce, whether machine learning is an appropriate approach, and how success should be evaluated. The responsibility may involve understanding practical requirements, examining available data, identifying prediction or generation targets, defining evaluation criteria, and comparing possible technical approaches.

Human skills are especially important for interpreting the real objective, recognizing unstated assumptions, understanding operational consequences, and deciding what trade-offs are acceptable. Agentic skills can complement this work by retrieving information, summarizing requirements, comparing technical approaches, examining relevant data characteristics, and organizing evidence.

For example, an organization may want to predict equipment failure. The engineer may first define what counts as a failure and how far in advance a useful prediction must be produced. An AI agent can then inspect available data descriptions, compare candidate modeling approaches, and identify possible target definitions. The engineer reviews these alternatives, determines whether the proposed targets correspond to real maintenance decisions, and may ask the agent to investigate another approach. Human and agentic skills are therefore activated iteratively until a suitable problem formulation is obtained.

1.4.2. Data Preparation and Feature Engineering

Data preparation and feature engineering ensures that the data used for model development appropriately represents the problem and can support reliable training and inference. The responsibility includes collecting, integrating, cleaning, transforming, labeling, sampling, validating, and representing data.

Agentic skills can contribute extensively because many data-preparation actions involve repeated inspection, comparison, transformation, and validation. Human skills remain essential for understanding the data-generating process, identifying inappropriate assumptions, recognizing leakage, and deciding whether a feature meaningfully represents the intended problem.

For example, an AI agent may profile a training dataset and identify a feature that has unusually strong predictive power. The engineer examines its meaning and discovers that the value is generated only after the outcome being predicted has already occurred. The engineer identifies the feature as leakage and instructs the agent to remove it, regenerate the dataset, and repeat the analysis. The collaboration therefore progresses through agentic inspection, human interpretation, and agentic re-execution.

1.4.3. Model Development, Training, and Experimentation

Model development, training, and experimentation involves selecting candidate methods, implementing models, configuring training procedures, testing hypotheses, comparing configurations, and determining which modeling direction should be pursued.

Agentic skills are well suited to repeated implementation and experimentation. AI agents can generate model code, create configurations, execute authorized training jobs, record outputs, compare results, and prepare experiment summaries. Human skills contribute to forming meaningful hypotheses, identifying experimental weaknesses, interpreting unexpected outcomes, and deciding how the modeling strategy should evolve.

For example, the engineer may hypothesize that a particular architecture will improve performance on a difficult subset of examples. An AI agent generates several implementations and runs the requested experiments. The engineer reviews the results and notices that the improvement appears only under one data split. The agent is then instructed to perform additional validation under alternative splits. The engineer interprets the new evidence and revises the modeling hypothesis. Human and agentic skills operate sequentially and iteratively throughout the experiment.

1.4.4. Model Evaluation, Validation, and Performance Optimization

Model evaluation, validation, and performance optimization determines whether a model performs adequately for its intended application and identifies where improvement is required. The responsibility includes measuring predictive quality, examining errors, testing robustness, comparing model versions, evaluating important data groups, measuring efficiency, and analyzing trade-offs.

Agentic skills can efficiently perform metric calculation, error grouping, robustness testing, segment comparison, and repeated evaluation. Human skills are necessary for deciding whether the measured differences are meaningful and whether an observed trade-off is acceptable within the application context.

For example, an AI agent may compare two candidate models and report that one has slightly higher average accuracy. The engineer examines group-level results and notices that the same model performs worse on a smaller group that is particularly important for the application. The engineer asks the agent to conduct additional error analysis and robustness tests for that group before deciding which model should proceed. Agentic measurement therefore supports, but does not replace, contextual engineering judgment.

1.4.5. Model Deployment, Serving, and System Integration

Model deployment, serving, and system integration moves a trained model into an operational environment where it can interact reliably with applications, services, data pipelines, and infrastructure. The responsibility includes packaging models, creating inference services, configuring environments, integrating interfaces, testing compatibility, and evaluating latency and resource use.

Agentic skills can support generation of deployment files, serving code, tests, configuration alternatives, and diagnostic information. Human skills are especially important when decisions affect architecture, reliability, security, resource allocation, or production consequences.

For example, an AI agent may generate two serving configurations and execute compatibility and performance tests. One configuration provides lower latency but requires substantially more computing resources. The engineer evaluates the production requirements and decides that the lower-resource configuration provides the more appropriate trade-off. The agent then prepares the selected configuration for further testing under the constraints established by the engineer.

1.4.6. Machine Learning Pipelines, Automation, and Lifecycle Engineering

Machine learning pipelines, automation, and lifecycle engineering creates repeatable workflows connecting data preparation, model training, evaluation, artifact management, deployment, monitoring, and retraining. The responsibility supports reproducibility and reliable movement of machine learning systems through different lifecycle stages.

Agentic skills can construct pipeline components, inspect dependencies, generate tests, execute workflows, and automate repeated operations. Human skills determine what should be automated, what validation conditions must be satisfied, and when human review or approval is required.

For example, an AI agent may execute a retraining workflow when new data becomes available. The resulting candidate model is automatically evaluated against predefined performance conditions. If the model satisfies those conditions, the engineer reviews important results and determines whether it should replace the current production model. If the conditions are not satisfied, the agent can perform additional analysis while the engineer determines the next corrective action.

1.4.7. Production Monitoring, Troubleshooting, and Continuous Model Improvement

Production monitoring, troubleshooting, and continuous model improvement maintains the quality and reliability of machine learning systems after deployment. The responsibility includes monitoring model performance, input distributions, latency, resource use, system events, and error rates; detecting changes; diagnosing problems; and selecting corrective actions.

Agentic skills are valuable because production systems can generate large quantities of logs, metrics, prediction records, and monitoring information. AI agents can continuously organize and compare this evidence, identify anomalies, correlate changes, and propose candidate explanations. Human skills are needed to distinguish correlation from causation, understand wider system dependencies, and determine the appropriate corrective response.

For example, an AI agent may detect that model performance declined shortly after an important input feature changed distribution. The engineer does not immediately assume that the distribution shift caused the problem. Instead, the engineer asks the agent to compare upstream data-processing changes, deployment events, and historical monitoring information. The agent returns additional evidence, and the engineer determines whether the appropriate response is retraining, rollback, data correction, or another system modification.

1.4.8. Machine Learning Software Engineering, Technical Coordination, and Knowledge Development

Machine learning engineering also requires reliable software, effective technical coordination, clear documentation, and continuous development of professional knowledge. This responsibility includes maintaining code quality, testing software, documenting experiments and assumptions, coordinating interfaces, communicating technical decisions, and evaluating new methods, tools, and libraries.

Agentic skills can support code generation, code review, test generation, documentation, experiment summarization, technical information retrieval, and comparison of new engineering approaches. Human skills are important for preserving system intent, evaluating integration consequences, coordinating across specialties, and deciding whether newly available technologies are appropriate.

For example, an AI agent may identify and summarize a newly released model-serving library and compare its capabilities with the existing serving system. The engineer evaluates the evidence in relation to current infrastructure, migration cost, reliability requirements, and team practices. The engineer may then ask the agent to construct a small compatibility test before deciding whether further adoption work is justified.

1.5. Changing Human and Agentic Involvement Across Machine Learning Engineering Responsibilities

Human skills and agentic skills are involved throughout machine learning engineering, but their relative involvement differs across responsibilities and may change during task execution.

These involvement levels should not be interpreted as permanent allocations. A single responsibility may contain actions in which human and agentic skills compete as alternative capabilities and other actions in which they collaborate sequentially, iteratively, or recursively.

The practical objective is to construct skill configurations that use the most suitable available capabilities for each part of the work and to revise those configurations when task performance evaluation shows that another arrangement could perform more effectively or reliably.

1.6. Working as a Machine Learning Engineer with Agentic AI

Agentic AI is becoming an increasingly powerful and promising part of machine learning engineering. Machine learning engineers should therefore be prepared to accept and embrace AI agents as complementary participants in professional workflows rather than treating them only as occasional tools for code generation or information retrieval.

This does not reduce the importance of machine learning engineering expertise. Human professionals remain responsible for the occupation and its corresponding responsibilities. They define objectives, understand professional and organizational context, determine acceptable outcomes, control consequential decisions, evaluate task performance, and decide how human and agentic skills should be configured and improved.

At the same time, agentic skills can greatly expand the amount and variety of work that can be performed. AI agents can inspect larger datasets, search broader technical information, execute more experiments, generate and test implementation alternatives, continuously analyze production evidence, and support repeated engineering actions at a scale that would be difficult to achieve through direct human effort alone.

The strongest machine learning engineering practice is therefore not based on maximizing human work or maximizing agentic work independently. It is based on finding appropriate configurations of human skills and agentic skills for the responsibilities, tasks, subtasks, and actions being performed.

Where a human skill and an agentic skill can independently perform the same action, skills competition can help identify which one is better suited to the current context. Where the work requires complementary capabilities, skills collaboration can activate human and agentic skills sequentially, iteratively, or recursively to achieve stronger task performance than either skill type could achieve alone.

Task performance evaluation then determines whether the selected skills and configurations satisfy predefined expectations. Successful skills can be retained and reused, while inadequate skills can be refined, reconfigured, or newly developed. The resulting capabilities return to an increasingly useful skill space that can support machine learning engineering and, where capabilities are reusable, other occupations as well.

For a machine learning engineer, working effectively with agentic AI therefore means treating AI as a powerful complementary capability that can be integrated into everyday professional routines while maintaining human control of occupational responsibilities. Through appropriate skill selection, competition, collaboration, evaluation, and continuous improvement, human professionals and agentic AI can work together to achieve machine learning engineering that is more efficient, scalable, adaptable, and reliable.