Denial

Matt Pocock’s Agentic Engineering Workflow (just copy him)

3 min read

Summary

  • Agentic Engineering Mindset – Treat AI as the utility that runs your development pipeline, not the primary builder of the product. The human remains the architect who defines problem boundaries, sets quality gates, and reviews work.
  • Phase‑Based Process – Break the journey into 7 clear phases:
    1. Research & Prototyping – Detect feasibility, gather requirements, and prototype prompts with the model.
    2. The Grill Session – Experiment, iterate and refine the prompt design using debugging from the model’s responses.
    3. PRD Writing – Translate the findings into a product‑ready PRD, including test cases and baselines for model output.
    4. Issue Slicing – Decompose the PRD into granular, AI‑applicable tasks and feed them to scheduler/agent engine.
    5. Implementation with AI Agents – Automate the code generation, docs, tests, and CI/CD configurations via the agent stack.
    6. Human‑in‑the‑Loop Review – Inspect each artifact; it’s the human who determines if the output is acceptable and safe.
    7. Deployment & Monitoring – Release the code, roll out feedback loops, and continuously fine‑tune the whole AI workflow.
  • Tooling & Architecture – Use a dedicated MCP (Management, Coordination, and Processing) stack, version control hooks, and a global memory file to keep the agent context.
  • Takeaway – Agentic engineering frees engineers from repetitive tasks but keeps the creative, architectural, and quality‐assurance parts entirely human. The model is a high‑velocity tool that accelerates and scales the engineer’s craft.

Key Ideas Explained

  • Human‑in‑the‑Loop – The final check guarantees ownership, skill preservation, and safety.
  • Phase granularity – Structured phases avoid scope creep and give the AI concrete boundaries.
  • MCP & Memory – Centralizing context prevents model drift and ensures consistent outcomes across tasks.
  • Parallel Agent Chains – Compute multiple agents in parallel to handle large‑scale projects more efficiently.
  • Feedback Loops – Every rollout feeds back into the prompt/artifact stack, continuously improving the agent’s performance.

Transcript

0:00 – “Introduction: The thesis of AI engineering – why designers need to become agent architects instead of just hand‑coding every line of code.” 4:20 – “Phase 1, Research & Prototyping: Explore model capabilities, create a small proof‑of‑concept, identify problems and define success metrics.” 12:45 – “Phase 2, The Grill Session: Debug prompts, modify the context, run lots of small trials and extract valuable patterns from model outputs.” 22:10 – “Phase 3, Writing the PRD: Convert the research into a product‑ready PRD, including detailed test cases and acceptance criteria.” 35:50 – “Phase 4, Slicing work into issues: Break the PRD into AI‑friendly, atomic tasks that can be fed to an agent scheduler.” 48:15 – “Phase 5, Implementation with AI Agents: Automate code generation, documentation, tests, and CI/CD pipelines via the agent stack.” 55:00? – “Phase 6, Human‑in‑the‑Loop Review: Inspect, validate, and approve each artifact.” ? — “Phase 7, Deployment & Monitoring: Release the code, monitor usage, capture feedback, and refine the entire agentic workflow.” ? – “Throughout: maintain a global memory file to keep context and avoid drift across long‑running sessions.” ? – “Throughout: build and orchestrate multiple agents in parallel for large projects, scaling from a single linear chain to a full policy‑shaped HPC model.”

Notes

  • AI give a very big boost to seniors, companies are not as incentives to hire juniors
    • skills are more important, AI is a multiplier
  • Pocock took his experience from teaching to write a skill to learn and dive deep
    • teach

Graph View

Backlinks