AI Agents as Janitors: A Minimum Functional Platform for the SDLC

AI Agents as Janitors: A Minimum Functional Platform for the SDLC

Executive Summary

This is a proposal for a shared, event-driven “janitor” automation service that helps engineering teams keep repositories clean, consistent, and compliant—without changing how features are built or shipped.

The system responds only to explicit triggers, makes small, bounded changes, validates them through existing CI pipelines, and proposes pull requests for human approval.

It is not autonomous development—it is disciplined cleanup at scale.

The Problem We’re Actually Solving

Modern engineering teams don’t struggle because they can’t write features.

They struggle because maintenance work piles up faster than humans can handle it.

Across large organizations, teams repeatedly lose time to:

• dependency upgrades

• repetitive refactors

• enforcing standards by hand

• policy drift across repositories

• low-value cleanup work done by senior engineers

These tasks are necessary, but they don’t differentiate the business—and they distract engineers from solving real problems.

Why “Agent-First” Development Fails

The industry’s current narrative suggests AI agents should continuously improve codebases.

In practice, that leads to:

• endless PRs

• constant CI churn

• reviewer fatigue

• unstable release cadence

• teams afraid to ship because the system is always changing

This proposal explicitly rejects that model.

If an AI system is “always running,” it’s already wrong.

The Janitor Model (What Actually Works)

In this model, AI agents behave like janitors:

• they only work when asked

• they receive explicit instructions

• they operate within strict boundaries

• they stop when the task is complete

No curiosity.

No creative rewrites.

No “improvements” that weren’t requested.

How the System Works (High Level)

1. An event occurs

   
   

   • a Jira ticket reaches “Ready”

   • a security or policy alert fires

   • a dependency update is required

   • a maintenance window opens

2. 
   

3. Context is pulled safely

   
   

   • Jira, GitHub, and observability tools are accessed through MCP services

   • MCPs enforce least privilege and data filtering

   • agents never see credentials directly

4. 
   

5. A bounded plan is generated

   
   

   • exactly what files can change

   • exactly what tests must run

   • no scope expansion allowed

6. 
   

7. A small code change is proposed

   
   

   • patches are generated

   • existing CI pipelines validate the result

   • no pipelines are replaced or bypassed

8. 
   

9. Humans approve

   
   

   • a PR is opened with an evidence pack

   • engineers review and merge on their terms

10. 
    

Where the Intelligence Actually Lives

The intelligence is not in the agent.

It lives in:

• architectural rules

• quality gates

• organizational standards

• definitions of “done”

The agents simply apply those rules consistently.

This is why the system scales across:

• repositories

• teams

• programming languages

• organizational boundaries

What This Is Not

This platform does not:

• write features

• control deployments

• auto-merge code

• continuously refactor

• replace developers or architects

Those ideas sound impressive in demos and fail in production.

Why This Saves Money (Without Reducing Headcount)

Organizations don’t save money by replacing engineers.

They save money by:

• reducing rework

• minimizing review fatigue

• preventing standards drift

• keeping senior engineers focused on high-value design work

You don’t pay janitors like architects—but you also don’t ask architects to take out the trash.

The Minimum Functional Product (MFP)

A realistic first version includes:

• Event-driven triggers only

• One or two high-ROI cleanup tasks

• One repository type

• Centralized policy enforcement

• Human-approved PRs

• Full auditability

No moonshots.

No platform rewrite.

No disruption to the SDLC.

The Real End Game

The goal is not “AI-first development.”

The goal is:

• cleaner repositories

• fewer low-value tasks

• consistent standards

• healthier delivery cadence

AI agents don’t need to be heroes to be valuable.

They just need to quietly clean up, on schedule, and get out of the way.

What’s Next

The next step is to define a minimum functional implementation—tools, staffing, timeline, and cost—that a real company could build, operate, and trust as a shared service.

No hype.

No fantasies.

Just a practical platform improvement.