Overview
From Assistance to Autonomy
The five levels of Agentic Engineering describe how independently AI participates in software development.
| Level | Role of AI | Role of the human |
|---|---|---|
| 1Level 1 | answers questions in an LLM chat | provides context and transfers results manually |
| 2Level 2 | provides direct support in the development environment | writes code and controls every step |
| 3Level 3 | handles specific development tasks independently | defines tasks and reviews the implementation |
| 4Level 4 | autonomously handles large parts of the development process | defines goals and approves critical steps |
| 5Level 5 | implements business requirements fully autonomously | defines the business requirement but does not intervene in implementation |
The higher the level, the less direct human control is required. Development ranges from simple use of an LLM chat to fully autonomous implementation of business requirements.
Reine LLM Chat-Nutzung
At Level 1, AI is used exclusively through a chat interface.
The developer copies code, error messages, or requirements into the chat and receives explanations, suggestions, or code examples.
Typical requests include:
- "Explain this error message."
- "Write a Java method."
- "Create a unit test."
- "How can I improve this code?"
The AI has no direct access to the project, repository, files, or development environment.
The human must provide the required context manually, verify the response, and transfer the generated code into the project.
Examples: ChatGPT, Claude, Gemini, or self-hosted models in a standard chat.
AI-Assisted Development in the IDE
At Level 2, AI is integrated directly into the development environment.
It supports the developer while coding and can consider the currently open code and parts of the project context.
Typical functions include:
- automatic code completion,
- suggestions for entire methods,
- generation of individual tests,
- explanation or revision of selected code sections,
- generation of documentation,
- inline suggestions directly in the editor.
The human continues to write code and control every step. The AI mainly responds to the current context and does not yet take on extensive development tasks independently.
Examples: GitHub Copilot Autocomplete, JetBrains AI Assistant, or basic Cursor inline suggestions.
Agentic Execution of Specific Development Tasks
At Level 3, the AI receives a specific development task and implements it largely independently within the project.
A typical instruction might be:
Implement a password reset feature.
The agent can then:
- analyze the repository,
- identify relevant files,
- create an implementation plan,
- modify multiple files,
- create new files,
- run commands and tests,
- detect and correct errors.
The human describes the task, defines acceptance criteria, and then reviews the result.
The AI no longer supports only individual programming steps, but takes on a complete development task.
Examples: Codex, Claude Code, Cursor Agent, or JetBrains Junie.
Largely Autonomous Development with Human Approval
At Level 4, AI independently handles large parts of the entire development process.
A single agent or several specialized agents can take on different tasks, such as:
- analyze requirements,
- prioritize tasks,
- create architecture proposals,
- implement code,
- generate and run tests,
- perform security checks,
- create pull requests,
- perform code reviews,
- resolve CI/CD errors,
- update documentation,
- prepare deployments.
The human primarily defines goals, constraints, quality requirements, and boundaries.
Implementation is largely autonomous. Human review or approval is still required at important checkpoints.
Typical approval points include:
- fundamental architecture decisions,
- Pull Requests,
- production deployments,
- security-critical changes,
- decisions with particular business or economic significance.
At this level, the human primarily acts as client, reviewer, and decision-maker.
Example: Lovable in Agent Mode, which plans, implements, and verifies complex features largely autonomously while the human sets goals, evaluates results, and approves critical steps.
Fully Autonomous Implementation of Business Requirements
At Level 5, a human defines business requirements or a desired outcome.
Such a requirement might be:
Develop a feature that lets customers view and download their invoices and raise a query if they identify discrepancies.
Once this business requirement is submitted, implementation proceeds entirely without further human intervention.
The AI system handles the entire software development lifecycle:
The system can independently:
- fachliche analyze requirements,
- derive technical requirements,
- evaluate and prioritize tasks,
- make detailed technical and business decisions,
- design software architectures,
- develop code,
- perform tests and quality checks,
- identify security risks,
- deploy software to production,
- monitor applications,
- analyze and resolve production errors,
- perform necessary optimizations and further development.
The human provides the initial business requirement but no longer intervenes in planning, development, review, approval, deployment, or operations.
Level 5 therefore describes fully autonomous software development within a defined business objective.
Control System
The Importance of the Harness
As autonomy increases, a harness becomes increasingly important.
The term harness, derived from the verb to harness, refers here to a governance and control framework. In Agentic Engineering, it defines how an AI system is directed, constrained, and made reliably useful.
A harness is the technical and organizational framework within which an AI agent is allowed to operate. It ensures that the AI follows defined rules, tools, and quality requirements instead of acting without control.
A harness can include:
- Architecturevorgaben
- Coding Guidelines
- Automated tests
- Access rights and permissions
- Security rules
- Privacy and compliance requirements
- Business rules
- Quality checks
- Permitted tools
- Logging and traceability
- Defined approval processes
- Limits for changes and deployments
- Automated rollback and emergency mechanisms
While humans still direct AI at lower levels, this control is increasingly transferred to the harness at higher levels.
The harness is therefore not merely a technical environment, but the central control system for reliable and secure agentic software development.