The 4-Step Agentic Engineering Workflow I Use to Ship Real Apps | Ras Mic

Summary · 5 min read instead of 47 min

Ras Mic (Micky) is a full-stack engineer who, per the episode description, has taught 100K+ people how to build with AI. He runs a consultancy that builds apps and agents for other companies, and has built his own projects, including a custom payment system for his church and an app called Bezalaw.

Ras Mic argues that a software factory should package a developer's workflow, skills and domain knowledge rather than let agents run unattended, because one wrong early assumption can ruin the whole output. His agents.md file chains four steps: isolate each task in a Git worktree, build to a service layer architecture, prove the work with video, screenshots or numbers, and ship through a Greptile review loop that repeats until the PR scores five out of five. He still tests features himself and uses the agent's evidence to shorten manual review. On models, he prefers Astra for performance and security audits and computer use, says Fable writes better code and UI, and argues that state-of-the-art models are not needed for every task.

Minute by minute

1. 0:56

   A software factory is your workflow. Ras Mic says the common idea of a software factory is usually a slop machine. His definition is packaging your workflow, skills and domain knowledge into an agents.md file and a set of skills. He says it should be model and harness agnostic, though he names Astra as the best model right now and says Fable 5.1 is great at writing code.

2. 2:03

   Isolate every task in a worktree. His agents.md explains how he wants the agent to work, not how the codebase is structured, since models can read that themselves. For each new feature or bug, a new feature skill branches a Git worktree off main so several agents can work in parallel; he shows five Codex tabs running. The skill also warns that worktrees do not isolate shared resources such as dev server ports and databases.

3. 7:14

   Build to a service layer architecture. The build step calls his code structure skill, which tells the agent to separate an orchestration layer that owns business rules, state transitions and auth checks from a service layer of reusable functions. Ras Mic says this keeps code readable if he ever has to read it himself, and limits the duplicate functions agents tend to write.

4. 10:41

   Agents must prove bugs and fixes. An evidence-driven testing skill has cloud agents record video of a bug before fixing it and of the fix afterward, useful when product staff file Linear tickets assigned to an agent. A before-and-after skill he took from Vercel captures screenshots, or numbers for non-visual work; one Astra performance pass cut navigation from 850 milliseconds to 60. He says this pushes the agent to prove its solution works.

5. 15:47

   Astra needs reminders to follow skills. Ras Mic says Fable 5.1 and GPT 5.6 Sol run the workflow end to end without manual triggering, but Astra sometimes skips steps such as the Greptile loop and has to be reminded. Peter Yang finds Astra timid and unreliable with skills; Ras Mic agrees it sometimes stops partway and needs to be told to continue.

6. 16:45

   Humans still test every feature. Ras Mic says what sets his setup apart is that he tests and uses each feature himself, out of empathy for users. At his consultancy, developers on the team review the code. He says the agent's evidence is meant to build trust and make manual review, which he calls the worst part of a developer's job, as short as possible.

7. 21:19

   Greptile loops until five out of five. Greptile, a code review agent, gives each PR a confidence score, and Ras Mic says top models usually start at four out of five. The Greploop skill feeds the feedback back through the build and prove steps, pushes, and waits for a new review until the PR hits five out of five or reaches an iteration cap of five by default; his version uses 10. He wrote only the new feature, code structure and evidence-driven testing skills himself.

8. 26:19

   Verification exists because trust is low. Ras Mic says he does not yet trust agents with software other people use. For personal tools such as an image-sharing service and a screenshot app called Still, he skips the workflow and cares only that they work. Yang says overnight runs collapse after one wrong early assumption, and Ras Mic compares this to poor communication in human engineering orgs.

9. 28:17

   Talking through plans instead of specs. Ras Mic plans by talking to the agent through Wispr Flow for five to 10 minutes or more, because he thinks out loud. A discussion about replacing a paid email service with Cloudflare email took about 27 to 30 minutes before he settled on tracking and custom domains. The conversation history serves as the spec, and he asks for files like a setup.md in the repo when he needs to come back later.

10. 32:57

    Astra for audits, Fable for code. Ras Mic says Astra almost always finds speed gains in modern frameworks and runs security audits without being moved to a lesser model, which he says happens with Fable, routing him to Opus 5. An Astra audit of his church payment system caught a security issue he calls potentially detrimental. He finds Astra's UI weak without reference images and says Fable is better at UI and writes code more likely to pass human review.

11. 38:10

    State of the art is not needed everywhere. For tighter budgets, Ras Mic recommends sticking with Astra and Codex, calling OpenAI more generous with resets than Anthropic. He used the cheaper Luna for joint tax work, with Sol as a sub-agent checking it. He expects model routers to matter soon and says top models should mainly orchestrate and review, since big models can do dumb things on small tasks.

12. 40:43

    Scheduled security checks for client apps. For a client app involving government companies, his team runs a daily scheduled security check on Cursor Cloud agents. He is building his own agent workspace on a home lab of Linux and Mac machines and says security audits should be mandatory for production apps. He once found 3,000-something worktrees filling his Linux box's memory, and Yang suggests a cleanup cron job.

13. 43:06

    Harnesses may absorb parts of the workflow. Ras Mic expects his setup to last because it describes how he works, though better models may make the code structure skill unnecessary and harnesses may handle worktrees and loops. He tests each new model with no agents.md or skills first. Both warn that lab engineers give advice with unlimited tokens, and Ras Mic says he supports open-weight models that run locally.

Published Oct 11, 2026.