Maven - Hardcore Agentic Engineering for builders who ship by Greg Ceccarelli

Maven - Hardcore Agentic Engineering for builders who ship by Greg Ceccarelli

ABOUT THE COURSE:

Advanced Engineering of AI Agents is a practical course for developers, team leads, and technical founders who want to move from occasional experiments with AI coding agents to managed, verifiable, and production-ready development with agents.

The course is focused on a new engineering practice called Agentic Engineering. It centers not around an individual prompt, diff, or generated line of code, but on the goal that the agent must accomplish and a system of evidence that confirms the accuracy of the result.

AI agents are already capable of writing code, exploring repositories, fixing bugs, and assembling features. But the real engineering challenge begins after the first prototype: how to ensure the agent has indeed completed the task, hasn't broken the system, hasn't forged the verification, and has delivered a result that can be maintained for months.

In 3 weeks, you will build your own agentic harness: from setting verifiable goals and managing the agent in the process to independent verification gates, stable context, and safe orchestration of multiple agents. The final result will be presented on Demo Day.

What You Will Learn in the Course

Design Goals for Autonomous AI Agent Work

You will learn to formulate tasks so that the agent doesn't just "try to do something" but moves toward a specific, verifiable, and limited result.

• Formulate limited goals with clear completion criteria.

• Turn natural language description of the result into automatic checks.

• Understand the difference between a simple prompt, goal-based workflow, and a full-fledged goal + rider.

• Create your own Agent Skills to turn Goal Engineering into a repeatable process.

• Distinguish situations where tests suffice and situations that require additional model-based assessments.

Verify Results Without Trusting the Agent's Words

An agent may confidently claim that a task is completed. In engineering systems, that's not enough: the result must be independently verified.

• Verify the agent's claims through an actually executed binary, actual exit code, and real data.

• Build separate verification gates with secrets that the agent itself does not have access to.

• Protect against counterfeit markers of successful verification.

• Use read-only executions to assess the agent's ability to identify real problems before the correction phase.

• Create workflows where "done" signifies a reproducible verifiable result, not a subjective feeling.

Manage the Agent During Task Execution

Instead of an endless stream of tokens and manually fixing the consequences, you will master the role of process operator: a person who timely stops, redirects, and controls the agent's work.

• Stop a faulty execution immediately after the first erroneous action.

• Redirect a stuck agent with specific facts, working code, and precise Git ranges.

• Determine that the agent is moving in the wrong direction before it spends a significant token budget.

• Analyze session history and your own intervention frequency.

• Develop the skill of timely intervention instead of micromanaging every step.

Design a Resilient Context for Long-term Work

Context should not live only in chat. You will learn to create an information environment where the agent can resume work after a restart, window change, or task transfer to another agent.

• Create hand-off documents where every important statement is linked to a specific file:line or reproducible command.

• Delegate complex research to auxiliary agents and receive only the necessary summary from them.

• Maintain an up-to-date AS-BUILT map for subsystems and components.

• Store long-term state in files rather than in chat history.

• Design the process so that an agent's failure does not lead to the loss of accumulated knowledge.

Orchestrate Multiple AI Agents

Different coding agents handle research, code generation, refactoring, and verification differently. In the course, you will learn to choose the right work scheme for the specific result.

• Identify for which tasks Claude Code, Codex, Cursor, and other AI coding agents are best suited.

• Choose between single execution, goal-based workflow, and a team of agents.

• Run agents in separate Git worktrees in parallel without branch conflicts.

• Understand when parallel work accelerates development and when it creates additional cognitive load.

• Apply semantic rebase: remove a tangled branch and regenerate changes from a clean base, preserving lessons learned.

Practical Course Outcome

By the end of the training, you will have not just a set of prompts, but your own agentic harness end-to-end, which can be applied in a real repository and adapted to the team's workflows.

• You will be able to set clear and verifiable goals for the AI agent.

• Learn to manage the agent's work in real-time.

• Independently verify results without relying on the model's statements.

• Build verification gates for the safe release of changes.

• Organize context and state for long-term development with agents.

• Learn to use multiple agents in parallel safely.

• Understand when to intervene and when it's better to let the agent continue working.

• Apply Agentic Engineering approaches in your own project.

The main skill of the course is to transition from the role of a person who writes code with AI to the role of an engineer who designs the system, directs agents, and proves the correctness of their created result.

Who This Course Is For

For Experienced Developers

The course is suitable for engineers who already use AI coding agents and want to go beyond prompt engineering and context engineering. You will learn to set tasks for agents, control execution, verify results, and maintain full control over changes in the code base.

For Technical Founders and Product Creators

If you actively use vibe coding but find the development speed unpredictable, the course will help turn working with AI agents into a systematic and repeatable process that can be sustained over months.

For Product Managers, Tech Leads, and Engineering Leads

The course will benefit specialists whose work increasingly focuses on task setting, specification writing, and result control. You will learn to design goals such that the agent can achieve them, and the result can be objectively verified.

Requirements for Participants

Engineering Experience

This is not a beginner course. You should be able to develop, test, and release software. The program is not an introduction to programming or artificial intelligence—it assumes professional engineering judgment.

Experience with AI Coding Agents

Regular experience using at least one AI coding agent is required, such as Claude Code, Codex, Cursor, or a similar tool. The basics of these tools are not covered in the course: the program focuses on task setting, agent management, and result verification.

Confident Use of Git and Testing

You should be fluent in working with tests, commits, branches, and Git worktrees. This will allow you to focus on Agentic Engineering practices rather than basic development tool setup.

Why It Matters for Production Development

Quickly generating code is no longer enough. In production, reproducibility, verifiability, safety, and the ability to maintain changes after the agent finishes a session are important. That's why the course emphasizes controlled autonomy: the agent is given the freedom to do the work, but the result undergoes independent engineering verification.

You will learn to use AI agents not as an unpredictable code generator but as part of a carefully thought-out engineering system, where every goal, verification, session, and context transfer has a clear role.

View File