AI coding agents – an exciting new era in software development
Understandably, the first encounter with powerful AI coding agents can cause some excitement and concern. Rapid advances in this field, particularly in ‘agent-based coding’ and local LLMs, are indeed impressive.
Experience and concerns at a glance
Can AI replace humans? A question that preoccupies millions. Well, not in the way some fear – at least not yet.
Expansion, not replacement
AI coding agents can be viewed as powerful tools for enhancing personal capabilities. For example, they can take over repetitive tasks, suggest code, and even debug, refactor or generate entire functions or classes. This leaves more time to focus on higher-level design, complex problem solving and creative aspects of software development.
Human control
Even the most advanced agents require human guidance, review and correction. There are sources of error, such as the generation of suboptimal code or misinterpreted requirements. Personal expertise about the overall system, business logic and potential edge cases remains indispensable.
Adaptation
The skills required in software development are evolving. While memorising code is becoming less important, the need for critical thinking, architectural design, understanding user needs and managing complex systems will continue to grow. Similarly, there is a growing demand for developers who are proficient in these AI tools.
New role distribution
Just as previous technological changes have given rise to new industries and job profiles, AI is likely to lead to new specialisations within programming. Examples include ‘AI prompt engineers’ for developers, ‘AI agent managers’ or specialists in the validation and integration of AI-generated code.
Experience with programming agents and local LLMs
It is both encouraging and rewarding to experiment and find tools that suit your workflow actively.
VibeCoding and Agentic Coding
Exploring Trae, Cursor, and Junie from JetBrains (especially if you already use the JetBrains ecosystem) is a smart move. Agentic Coding, where AI tasks can be planned, executed and iterated, represents a significant leap beyond simple code completion.
Local LLMs for cost savings
The primary motivation for using local LLM models to avoid high API costs from OpenAI, Anthropic, and other providers for private projects is a common concern in practice.
Cost versus performance
Operating local LLMs has advantages and disadvantages. API costs can be avoided, but other costs arise:
Hardware
Powerful GPUs (such as consumer cards from NVIDIA or even more robust professional cards) are often necessary to achieve reasonable inference speeds with larger models.
Energy
Operating GPUs requires a significant amount of electricity.
Time and effort
Setting up and managing local LLM environments can be time-consuming and require technical expertise.
Performance Local models may not consistently achieve the peak performance or extensive context windows of the most powerful cloud-based models, especially for very complex tasks. However, they are constantly improving at solving a wide range of everyday coding tasks.
Assessment
The cost of operating a local LLM depends on the following factors:
Larger models (e.g. seventy billion parameters versus seven billion parameters) require more VRAM and computing power.
Inference speed: Speed of feedback
Usage: Hours per day of LLM application
Electricity tariff: Varies by region.
Hardware depreciation: Cost of the GPU over its entire ‘lifespan’.
For personal planning, the power consumption of a GPU such as an NVIDIA RTX 4090 (approximately 450 W at full load) can be researched and calculated based on local electricity rates. For example, if you run it for eight hours a day, twenty days a month, at an average of 300 W and an electricity price of £0.30, the monthly cost is around £14.00.
0.3 kW × 8 hours/day × 20 days/month × £0.30/kWh ≈ £14.40/month
This is intended to give you an idea of the operating costs you may encounter. Initial hardware costs are not factored into the calculation.
Conclusion
These tools are helpful and are changing the landscape. Nevertheless, human creativity, personal problem-solving skills and contextual understanding will remain indispensable. Those who experiment, learn, adapt and are willing to change will remain at the forefront of this exciting new era of software development.