§ AGENTIC ENGINEERING
I architect it and own what ships. AI writes much of the code, under a review process I built.
I use Claude Code every day and hold its output to the same independent-review standard as my own. Here are the practices that make that work, and the automations I’ve built to run them.
- numbered decisions
- 320+
- test code to app code
- ~1:1
- scheduled AI routines
- 5
- unapproved merges
- 0
The weekly loop
FIG. A-01 · THE LOOP AS A DISTRICT
- COMMAND CENTRE · Launch headless AI runs and watch the results stream back.
DATA FLOW
- RETROwhat shipped, what broke
- SCOPED TASKwritten down first
- BUILDone session, one branch
- REVIEWindependent
- HUMAN MERGEmy approval
- LIVEmonitored daily
I tried five AI agents in parallel and couldn’t keep track of them. One session per task now
- numbered decisions
- 320+
- test code to app code
- ~1:1
- unapproved merges
- 0
WHAT IT IS
A weekly plan, build and review cycle for my products.
HOW IT WORKS
A retro feeds a queue of scoped tasks. One AI session builds each on its own branch, an independent reviewer checks it, and I approve every merge and deploy.
WHY IT MATTERS
AI output is fast but unaccountable on its own. The loop puts a written plan before the code and a human after it.
Scoped tasks
FIG. A-02 · A TASK AS A STREET
DATA FLOW
- RETROweekly
- TASK QUEUEordered
- GOALone line
- DONE CHECKhow we know
- LEFT OUTlisted
WHAT IT IS
Every task is written down before an agent touches it.
HOW IT WORKS
One line on what the task delivers, the check that will show it’s done, and a list of what it leaves out.
WHY IT MATTERS
Agents drift. Listing what a task leaves out is the cheapest way to stop it growing while nobody is watching.
Independent review
FIG. A-03 · A REVIEW AS A STREET
DATA FLOW
- BUILDits own branch
- DIFF ONLYno build history
- REVIEWERseparate session
- FINDINGSranked by severity
- FINAL SWEEPdeleted or overwritten data
Review goes deeper on code that can delete or overwrite data
WHAT IT IS
A separate AI session reviews each task without seeing how it was written, and goes deeper where the code can delete or overwrite data.
HOW IT WORKS
Findings are ranked by severity, review rounds are capped, and a last check looks for anything that deletes or overwrites data.
WHY IT MATTERS
A reviewer that watched the build inherits its blind spots. One that didn’t has to read the code as it is.
The decision log
FIG. A-04 · A RULE AS A STREET
DATA FLOW
- INCIDENTwhat went wrong
- MEASUREDcause confirmed
- DECISIONnumbered
- RULEin the process
- PRUNEDwhen outgrown
- numbered decisions
- 320+
WHAT IT IS
A numbered log of engineering decisions, 320+ of them on Kiwi Game Bazaar.
HOW IT WORKS
New rules are only added after a measured incident, and the log records why. Rules get pruned when the process outgrows the product.
WHY IT MATTERS
A new developer, or a new agent, can check why a rule exists before changing it.
Command Centre
FIG. A-05 · THE DASHBOARD AS A STREET
DATA FLOW
- DASHBOARDlaunches a run
- HORIZONqueued as a job
- HEADLESS RUNClaude Code
- REVERBstreamed live
- DASHBOARDresults + history
WHAT IT IS
An internal Laravel/Inertia dashboard over my Claude Code setup: its skills, scheduled routines and memory.
HOW IT WORKS
It launches headless AI runs through Horizon queues and streams the results back to the browser over Reverb.
WHY IT MATTERS
AI runs started and watched from one screen instead of a pile of terminals, with results streamed back live.
Scheduled routines
FIG. A-06 · A WEEK AS A STREET
DATA FLOW
- PLANNINGweekly
- DEV SESSIONdaily
- STATUS REPORTdaily
- HEALTH CHECKdaily · Nightwatch MCP
- AUDITweekly
- planning retro
- Weekly
- dev session + status report
- Daily
- site-health check
- Daily
- security + test audit
- Weekly
WHAT IT IS
Five scheduled routines on a live product.
HOW IT WORKS
A weekly planning retro, a daily development session and status report, a daily site-health check through a Nightwatch MCP connector, and a weekly security and test-quality audit.
WHY IT MATTERS
The product stays monitored and its backlog keeps moving, and nothing they produce merges without my approval.
MCP integrations
FIG. A-07 · AGENT TOOLS AS A STREET
DATA FLOW
- CLAUDEAI agent
- MCP SERVERtyped, guarded tools
- ADMIN ACTIONSConsoleCore
- NIGHTWATCHproduction health
WHAT IT IS
An MCP server on a live tournament platform, and a Nightwatch MCP connector for production health.
HOW IT WORKS
ConsoleCore’s admin actions are exposed as tools on the MCP server, and Nightwatch is wired in as an MCP connector.
WHY IT MATTERS
Agents act through named tools instead of a shell, and merges still need me.
A wiki that maintains itself
FIG. A-08 · KNOWLEDGE AS A STREET
DATA FLOW
- COMMITany tracked repo
- GRAPH HOOKcode graph rebuilt
- WIKIkept current by AI
- NEXT SESSIONstarts with context
WHAT IT IS
An Obsidian knowledge base that AI keeps current.
HOW IT WORKS
Commit hooks mirror each repo’s code graph into the vault, and AI keeps the wiki current from there.
WHY IT MATTERS
The next session, human or agent, starts from what’s already known instead of working it out again.
§ AT A GLANCE · CLICK TO OPEN







