Reviewable ADRs, AI by default, and weekly readings! 💡

2025-04-30 · Source: Refactoring · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering, Robotics & Autonomous Systems · Depth: Intermediate, short

Summary

This intelligence brief highlights several key developments and insights for technical professionals. It introduces Kestra, an open-source platform for scaling AI agents into deterministic and observable workflows, noting its use in replacing cron jobs. A significant focus is on Architecture Decision Records (ADRs) as reviewable artifacts in AI coding, exemplified by Tolaria's 120+ ADRs, which help steer agent behavior and capture design judgments. The brief also emphasizes adopting AI "by default" to accelerate learning its failure modes, a strategy advocated by Stuart Caborn of loveholidays, where teams intentionally use AI even when manual coding might be faster to build crucial feedback loops. Additionally, it briefly covers readings on using sensors for coding agents, the cost inefficiency of running LLMs locally on Apple Silicon compared to services like OpenRouter, and how LLMs are diminishing programming language lock-in.

Key takeaway

For AI Engineers and MLOps teams scaling agentic workflows, consider implementing Architecture Decision Records (ADRs) to formalize design choices and provide reviewable artifacts, improving agent consistency and future work. Additionally, adopt an "AI by default" strategy within your team; intentionally using AI for tasks, even when slower, will rapidly expose its failure modes and accelerate your collective learning curve. This approach builds essential feedback loops, making your AI systems more robust.

Key insights

ADRs and "AI by default" accelerate learning and control in AI development workflows.

Principles

• ADRs make AI design judgments reviewable.
• Embrace AI by default to learn its failure modes.
• Sensors provide automated guardrails for coding agents.

Method

Document AI design choices via immutable ADRs, and integrate AI into daily tasks by default to surface and learn from failures.

In practice

• Implement ADRs for AI agent design.
• Use Kestra for AI workflow orchestration.
• Mandate AI use before human support.

Topics

• AI Workflow Orchestration
• Architecture Decision Records
• AI Coding Agents
• LLM Cost Optimization
• Feedback Loops
• Static Analysis Sensors

Code references

• refactoringhq/tolaria

Best for: AI Architect, AI Engineer, MLOps Engineer, Director of AI/ML

Related on AIssential

• From Workflows to Multi-Agent Systems: How to Choose — Choosing between workflows, single agents, and multi-agent systems depends on task complexity, dynamism, and cost-benefit analysis.
• Decentralizing Architectural Decisions with the Architecture Advice Process — Decentralizing architectural decisions via an advice process empowers teams and distributes accountability.
• The Sequence Opinion #860: Every Company’s Last eXam: Some Reflection About Practical AI Evals — Company-specific, dynamic evaluations are now essential for production AI, forming a fourth pillar alongside compute, data, and models.
• LAI #115: The Hidden Cost of “Agent-First” Thinking — AI system failures often stem from architectural and reproducibility issues, not just model performance.
• stop trying to build an ai employee right now — Focus on small, reviewable "packets" of work for initial AI automation wins, not giant autonomous agents.
• The Agentic Shift: Why Engineers Are Becoming Workflow Architects — Agentic AI transforms engineering by autonomously planning and executing tasks, requiring human oversight for safety.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Refactoring.