Fragments: July 13

2026-07-13 · Source: Martin Fowler · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering, Robotics & Autonomous Systems · Depth: Advanced, long

Summary

The July 13 "Fragments" from a Thoughtworks retreat highlight key trends in software development with AI. Discussions covered "Harness Engineering," focusing on context management for LLMs to improve focus and computational sensors using languages like Rust and property-based testing. Self-hosting open-weight models is gaining traction due to increasing token costs, a desire for sovereignty, and information security concerns, despite talent and infrastructure challenges. Kief Morris's narrative suggests a core debate across sessions: how much leeway to give agents and how to maintain confidence in their actions. Sam Ruby's "Bring Me a Rock" session re-frames iterative LLM interaction as managing by objective, emphasizing human judgment for acceptance criteria. Other notes include local models for programming (Qwen 3.6), cost-saving tips with Anthropic Fable, AI's impact on developer education, and the future of cross-platform UIs.

Key takeaway

For MLOps Engineers evaluating LLM deployment strategies, prioritize self-hosting open-weight models to gain sovereignty and control costs, especially with increasing token prices. Focus on managing agents by objective, defining clear acceptance criteria, and investing in "Harness Engineering" to ensure model reliability and efficient resource use. This approach mitigates risks associated with external dependencies and unstated objectives.

Key insights

The core challenge with AI agents is balancing their autonomy with human confidence in their actions.

Principles

• Harness Engineering improves LLM focus and reduces token usage.
• Manage LLMs by objective, not by method.
• Human judgment is crucial for acceptance criteria.

Method

Context management for LLMs involves limiting `agents.md` files to under 200 lines. Computational sensors benefit from shifting to Rust and using property-based testing. Teach models to pick optimal models for tasks.

In practice

• Consider self-hosting open-weight models like Qwen 3.6 for cost savings.
• Fine-tune models for domain-specific tasks to reduce token consumption.
• Use Anthropic Fable's judgment to delegate smaller tasks to cheaper models.

Topics

• Harness Engineering
• LLM Context Management
• Self-Hosted Models
• Open-Weight Models
• Agentic Programming
• Objective-Based Management

Best for: CTO, VP of Engineering/Data, Director of AI/ML, AI Engineer, Machine Learning Engineer, MLOps Engineer

Related on AIssential

• Harness Engineering — Harness engineering builds reliable LLM systems by managing the operational environment around the model.
• 🥇Top AI Papers of the Week — LLM internal states, external environments, and architectural choices profoundly impact performance, safety, and cost.
• not much happened today — AI development is converging on agentic platforms, optimized inference, and cost-effective open models, shifting focus from raw model intelligence to system-level engineering.
• YES: Harness Self-optimization w/ 9B LLM (Local AI) — Small LLMs can update agent harnesses effectively, but solver model capability dictates the benefit from these updates.
• The Model Is the Easy Part: Why Every AI Agent Needs a Harness — The model is the brain; the harness is the body, enabling LLMs to act in the real world.
• LAI #134: Your First LLM App on AWS for Under a Dollar — The newsletter highlights practical AI deployment, robust agent evaluation, and efficient LLM application development.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

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