Vibe coding was asking CGPT or Claude for code then implementing it yourself.
Agentic coding is where you give a prompt to an “agent” like ClaudeCode or Codex and it will create the right folders, create the codes in the right places, test the code itself in the command line, then view the bugs itself and fix them iteratively. These agents can run for hours fixing bugs while you crack on with more important things.
Many Agents are not for Coding, but instead for “workflows” – automating many manual processes on a PC, e.g. performing web scraping, web searches, or email marketing.
my tips on Agentic (and sub-agentic) Coding
Safety: Create a Sandbox – Agents can’t really be contained and have been known to delete entire projects. Use Docker to create a wrapper around the Agents activities. Or make a new user account on your PC with non-Admin privileges and let Microsoft file permission create a boundary. Use the “–yes” command only if the Agent can’t access non-backed-up files – Take a Belt & Braces approach. Create code backups at least daily – in case the Agent deletes everything
Your Role: As someone progressing Agentic Coding your journey begins as (i) code reviewer, then (i) you start employing sub-agents to discuss/check the work of the main agent for you, (iii) then you build loops where the code is iterated on and bugs are worked on with a team of sub-agents agents working autonomously: testing/deploying/measuring/fixing bugs and repeat.
I created sub-Agents, Karen , Edd (looks for edge cases) and Mentor (concerned with long-term stability) to challenge and review the main Agents solutions and final work
The Agentic Coder becomes a master of the levers & limitations of the Agents. For example:
Problem: Agents cannot check their own work well (Solution: employ a subagent to do this).
Problem: Agents can get confused if their starting memory file (CLAUDE.md) gets too large or contradicts itself (Solution: frequently amend/shorten/clarify your .md file) Whenever an agent does something wrong I ask: what line in the CLAUDE.md file made you do X?
Problem: Agents keep lying and saying they checked everything. Solution: Enforce that Agents provide evidence for every claim made in the CLAUDE.md (A tip here is you must specify logs cannot be used – as these can be stale with code changing every minute)
In an automated coding system which runs for hours, I created a “Compliance” Agent to sign off the main agents work
Creating codes from scratch? you will need to pass as much context to the Agent as possible, detailed plans and examples of what you want. Get the Agent to describe the problem back to you before it starts. (An advanced concept is to use an MCP like Playwright which allows the Agent to see the screen and edit UI more intuitively). What some people are doing is creating a closed off environment then allowing Claude the hours it may need to create the software.
The CLAUDE.md file is the “initialiser” file that all agents read before a session starts, it sets the scene for the agents behaviours and limitations and what it will optimise for
Critical questions to ask when things go wrong: What in the CLAUDE.md made you make X mistake? —> then update the CLAUDE.md file appropriately
Personally I like to work with ClaudeCode but have Codex (OpenAI) as one tab and I ask it questions about my code base sometimes then pass it to Claude – the insights are always completely different to Claude!
Migrations (e.g. Java to Kotlin) are relatively smooth because you provide the Agent with the old project to refer to. And use “–yes” to allow it to work for hours (with sub-agents critiquing and checking the work)
Github Mac Test Automation – Especially useful if you don’t have a Mac, but want to deploy on Apple products. You set Github as a connected experience on the Claude website. Then you ask Claude to login to Github via CLI. Once inside you ask Claude to push the latest update, then read the compilation/error logs, then make changes (sub-agents review) then push the code again and iterate as many times as you wish.
For Java builds on Windows I like to run the following –> Agent discusses fix with sub Agents –> test with gradle –> auto test with headless emulator (a phone in the PC) –> check for errors and iterate from beginning
I set a confidence threshold for completing tasks, because Claude kept rushing broken code, this is a great way to improve accuracy
William Gillbert has over fourty roles in his repertoire
01.
Elit pharetra dignissim
Vitae mus aliquam massa adipiscing
02.
Etiam ipsum volutpat
Tincidunt fringilla fringilla nibh gravida
03.
At arcu blandit turpis
Vitae et nunc eu
04.
Turpis cras turpis velit
In arcu fermentum
05.
Viverra nunc eleifend cursus
Vitae mus aliquam massa adipiscing
06.
Habitant aliquet turpis gravida
Tincidunt fringilla fringilla nibh gravida
07.
Porttitor mattis orci morbi
Vitae et nunc eu
08.
Enim egestas dui amet
In arcu fermentum
09.
Commodo arcu integer
Vitae mus aliquam massa adipiscing
10.
Platea viverra varius auctor porta
Tincidunt fringilla fringilla nibh gravida
11.
Odio volutpat varius maecenas
Vitae et nunc eu
12.
Porttitor in habitant
In arcu fermentum
Upcoming Show
Risus etiam dui
Ultricies nibh morbi amet fames in enim turpis aenean neque commodo nisi, tristique et nulla tellus.
