Quick Summary: AI Coding Agents Help, But There’s a Catch
James Shore, a veteran software developer, wrote a compelling analysis arguing that AI coding agents that only increase code output without reducing long-term maintenance costs will become traps that tank productivity below pre-AI levels within months.
The principle is simple: if AI helps you write code 2x faster, the resulting code must cost half as much to maintain. Otherwise, code volume increases but quality doesn’t follow, and your team drowns in bug fixes instead.
Before investing in an AI agent, teams need to ask: “Is the code AI writes easier or harder to maintain than what we write ourselves?” If harder, you’re just buying temporary speed.
Why Maintenance Cost Matters More Than Code Output
Shore estimates from industry experience that code written in a single month generates roughly 10 days of maintenance work in its first year, then about 5 days per year going forward.
This means the faster you write code, the faster maintenance burden accumulates proportionally. If AI-generated code has high maintenance costs, team productivity drops below 50% within roughly 10 months.
But if the output is high quality with half the maintenance cost, productivity stays above 50% for up to 68 months. That’s the difference between AI that genuinely helps and AI that creates technical debt.
The Formula to Remember: Output and Maintenance Must Balance
Shore distills it into a simple formula:
“If you double your output, you need code that costs half as much to maintain. Three times as much code, one third the maintenance.”
In a scenario where AI doubles output but also doubles maintenance cost, the productivity benefit lasts only about 5 months before things get worse than before. Shore calls this “trading a temporary speed boost for permanent indenture.”
When Legacy Code Becomes a Major Problem
In real dev teams, the common scenario is legacy code with no documentation, functions hundreds of lines long, no test coverage, and fixing one bug reveals several more.
Maintenance costs eat a large proportion of dev time because every modification requires understanding the code first. Business logic is scattered with no clear patterns.
This is where AI coding agents have real potential — if they can read legacy code, explain how it works, and suggest safe refactoring approaches. But the critical question per Shore’s thesis is: does the AI-refactored code actually become easier to maintain?
AI Coding Agent Market Position
AI coding agents span a wide spectrum, from code completion to autonomous end-to-end agents. GitHub Copilot focuses on inline code suggestions, while tools like Cursor or Claude Code work as agentic systems that can manage entire codebases.
The interesting position is tools that emphasize maintenance and code quality over raw code generation, since enterprises need maintainability more than rapid development. As Shore points out, the problem is there’s no clear evidence yet that any AI tool actually reduces maintenance costs long-term.
Approach Comparison: Manual vs AI-Assisted
| Factor | Manual Coding | Using AI Coding Agent |
|---|---|---|
| Coding Speed | Based on dev skill | Faster (but needs review) |
| Maintenance Quality | Depends on team skill | No clear evidence yet |
| Learning Curve | Low | Medium — prompt/config learning needed |
| Cost | Dev salary | Salary + license + infrastructure |
| Tech Debt Risk | Controllable | High without proper review |
Shore warns that the evidence he found suggests most AI coding agents currently increase maintenance costs rather than reduce them, since generated code requires additional review and fixes.
What AI Agents Do Well Today
Boilerplate code — Generates CRUD, API endpoints, database models quickly, saving project setup time.
Code explanation — Reads legacy code and explains functionality, helping new developers understand codebases faster.
Bug detection — Scans for potentially problematic patterns like memory leaks or security vulnerabilities.
Test generation — Writes basic unit tests, though edge cases requiring business context understanding still need human authoring.
What they still struggle with is architecture decisions and business logic — the parts that impact maintenance cost the most.
Current AI Coding Tool Pricing
| Factor | GitHub Copilot Pro | Tabnine Code Assistant |
|---|---|---|
| Price/month | $10 | $39/user |
| Strength | Code completion + chat | Privacy-first, self-hostable |
| Auto Refactoring | Limited | Limited |
| Best for | General developers | Privacy-focused enterprises |
Pricing starts at $10–$39/month per user, not including infrastructure costs for self-hosted models or API calls exceeding quotas. Starting June 2026, GitHub Copilot will transition to usage-based billing via AI Credits.
Pros and Cons of Using AI Coding Agents
Pros
- +Significantly reduces time on boilerplate code and repetitive tasks
- +Helps explain legacy code, speeding up onboarding for new devs
- +Scans for basic bugs and security issues
- +Auto-generates documentation and unit tests
Cons
- −No evidence of long-term maintenance cost reduction yet (per Shore)
- −AI-written code needs review every time — context misunderstandings happen
- −Complex business logic still requires human decision-making
- −High upfront costs: license, training, and infrastructure
Hidden Costs
Beyond the license price, there are costs many teams forget. Training the team to use AI agents effectively requires reworking workflows and review processes entirely.
Infrastructure is another major category. Cloud-based means monthly API call fees; self-hosting requires GPU servers.
Customizing AI agents to match company coding standards also consumes additional budget. Someone needs to configure and maintain integration with existing tools.
Most importantly, per Shore’s framework: if maintenance costs don’t actually decrease, all these expenses become pure overhead with no ROI to offset them.
Who Should Try It, Who Should Wait
Should try: Teams with several developers where maintenance consumes a lot of time, especially legacy codebases needing frequent updates. But a strong review process must be in place first.
Should wait: Small teams with simple code, or teams without proper git workflow and code review processes — an AI agent may create more problems than it solves.
Conclusion
AI coding agents are powerful tools, but Shore’s warning is clear: don’t look only at coding speed. Maintenance cost is what matters. If AI-generated code is harder to maintain than human-written code, you’re just trading temporary speed for permanent technical debt.
Recommendation: Start with a small project. Measure both output speed and maintenance cost. If maintenance genuinely decreases, scale up. If not, don’t force it.