Worth your time?
My Agentic Engineering Workflow (step by step workflow) by Ras Mic
Watch on YouTube
This is primarily a teaching video: a practical, step-by-step agent engineering walkthrough with one clear sponsor read but strong demos and measurable results.
Why this score
- Scored from the transcript, summary, and output signals.
The short version
The creator builds an artifacts feature for an agent-first app called Pluto using an agentic engineering workflow: Cursor with GPT-5.5 (extra high, fast) for generation, Whisperflow for voice prompting, and Grapile with the /gp-loop skill for iterative code review and automated fixes. He demonstrates designing the feature, generating a plan, splitting the work into small PRs, running GP loop to iterate reviews to 5/5 confidence, and deploying to a staging branch while highlighting product and security trade-offs (sandboxing, version history, preview behavior).
What it teaches
- Treat agentic development as a controlled feedback loop: use automated reviewers (Grapile) plus an orchestration skill (/gp-loop) to close the review->fix->re-review cycle. Apply this when you want to scale code checks and reduce manual back-and-forth; avoid it for tiny one-off edits where human review is faster.
- Ship in minimal increments so automated reviewers work: split cross-cutting features into stacked PRs that each expose a single review surface (parser, persistence, UI). Use this when a feature touches multiple systems; do not use monolithic PRs because agents miss issues and the loop degrades.
- Use subagents for non-blocking research or long-running tasks: delegate fetching/aggregation to subagents so the main chat keeps responding. This is helpful during interactive product demos or parallel data collection; skip subagents for tightly coupled tasks that require single-threaded state.
Key moments
- 0:42Listing the core workflow tools (Cursor + GPT-5.5 extra high fast, Grapile, Whisperflow): sets the technical stack and cost/performance trade-offs used throughout the build
- 4:50GP loop diagram and behavior: explains how Grapile reviews, returns a score, and how the /gp-loop skill automates iterative fixes until the review target is met
- 7:02Why GPT-5.5 extra high fast: model choice justification, better at writing tests and handling complex feature implementation
- 12:08Plan generation complete: the agent produces a staged rollout plan (support markdown/HTML/SVG/code, preview sandboxing path) that becomes the development blueprint
- 24:18First automated review result (3/5): Grapile surfaces security and draft-leak issues, triggering the need for iterative fixes
- 25:27Fire /gp loop: the author runs /gp loop in Cursor so the agent reads comments, edits code, pushes commits, and triggers new reviews automatically
Skip the video. Keep the lesson.
More scored videos
