Agentic Engineering Workflow: what I took from the workshop

I attended Kyrylo Sulimovskyi’s Agentic Engineering Workflow workshop. Kyrylo is a Head of Engineering, and he showed his day-to-day process with agents on a real product.

I had already been using Claude Code and Cursor, but the results were inconsistent. I wanted to see a systematic way to define tasks, manage context, and verify the output so agent work becomes more predictable.

The format was two days, nine blocks, six hours of material, and three and a half hours of live Q&A.

Day 1: organizing agent work

• Context as a resource. Why important details get lost in long sessions, how context degrades, and when to start with a clean context.
• Repository setup. CLAUDE.md, path-specific rules, and levels of instructions that help an agent follow the project contract.
• Subagents and agent teams. Separate context for roles and a shared task list for coordination.
• The Ralph loop and model switching. The advisor pattern, OpusPlan, and verification after each iteration.

Day 2: the full product cycle

• Specify and clarify: from an idea to a PRD and acceptance criteria.
• Design: SAD, ADR, and C4 to record architecture decisions as the work progresses.
• Tasks and test plan: decomposition, dependencies, and a basis for TDD.
• Implementation: vertical slices, parallel agents, and verification gates before completion.

What I am taking into my practice

My main takeaway: agents need a process with checkpoints. A specification states the expected behavior, small slices limit risk, and tests and verification show whether the work is actually done. Code produced quickly is not finished engineering work until its result has been checked.

For people who write code every day and want more consistent work with agents, this workshop offered a useful process to adapt to their own repositories.

View my Agentic Engineering Workflow certificate.