Fragments: July 6
Summary
The recent Future of Software Development Retreat in Europe revealed a significant shift in agentic development, moving from aspirational discussions to practical, production-level implementation. Attendees noted the emergence of "harness engineering" and a new focus on managing escalating token costs, a stark contrast to earlier efforts to incentivize AI interaction. A key debate centered on whether traditional software architecture remains vital, with one view suggesting AI's "Galaxy Brain" negates its need, while another, supported by Laura Tacho, posits that good modularity and naming benefit AI agents as much as humans. This concern over token costs is echoed by leaked reports from companies like Citi and Amazon, showing bills rising from \$5 million to \$15 million in months, prompting efforts to throttle AI use and even make models "speak like cavemen" to save costs. Separately, a fraudulent DMCA claim against Gergely Orosz's article highlights risks in digital content moderation.
Key takeaway
For AI Engineers and Software Architects integrating agentic development, prioritize robust software architecture that enhances both human and AI comprehension, as this directly impacts long-term maintainability and token efficiency. Proactively manage escalating token costs by optimizing AI usage and exploring cost-reduction techniques like simplified prompts. Integrate AI-driven quality checks and documentation generation into your daily workflows to leverage agents effectively while maintaining human oversight and accountability.
Key insights
Agentic development is moving into production, shifting focus to practical challenges like architecture, token costs, and ethical engagement.
Principles
- Good architecture, emphasizing ease of change, benefits both human and AI agents.
- Token costs are a rapidly escalating concern requiring active management.
- Engaging with AI's ethical challenges is more effective than renouncing its use.
Method
A proposed agentic workflow involves agent collaboration, Architectural Decision Records (ADRs), task list generation, agent completion, and automated quality checks.
In practice
- Measure design quality by observing token cost changes for modifications.
- Develop abstractions to improve communication with AI agents.
- Implement overnight AI-driven quality checks for human review.
Topics
- Agentic Development
- Software Architecture
- LLM Costs
- Harness Engineering
- AI Ethics
- Digital Rights Management
Code references
Best for: CTO, VP of Engineering/Data, Director of AI/ML, Software Engineer, AI Engineer, AI Architect
Related on AIssential
Editorial summary, takeaway, and curation by AIssential. Original article published by Martin Fowler.