The Frontier

Dex Horthy defines context engineering as deabstracting tools like RAG and memory to focus on crafting token inputs and structured outputs for higher-quality AI applications.

Horthy argues frontier LLMs follow 150-250 instructions before performance degrades, a limit grounded in transformer architecture fundamentals.

He splits context engineering into information budget and instruction budget, where conflicting or distant instructions in the window degrade model reliability.

Horthy observes that the smart zone of a context window is roughly the first 100k-200k tokens for frontier models, beyond which output quality often declines.

Intentional compaction, creating concise artifacts like research summaries to reset context windows, is a core tactic for maintaining model performance in complex tasks.

He warns that a model replying 'You're absolutely right' after a correction often signals a degraded trajectory, suggesting starting a fresh session.

Loop engineering automates verification by letting models self-correct against deterministic checks like unit tests or linters, but Horthy cautions it generates overwhelming code volume.

Horthy advocates for slow loops - cron-triggered agents that fix one issue nightly - as a sustainable way to improve codebases incrementally without losing architectural control.

He built a lights-off software factory in July 2025 and shut it down by November after unmaintainable code accumulated, concluding AI lacks intuition for long-term software architecture.

Horthy notes current coding benchmarks like SWE-bench reward functional fixes but fail to evaluate maintainability, which manifests as a cost months later in deteriorating codebases.

He defines a dark factory as a fully automated software production line with no human input, analogous to robotic car factories, and warns it risks existential breakdowns.

Horthy frames token harder as maximizing raw AI utilization, while token smarter focuses on human-in-the-loop leverage for maintainable, architecturally sound outputs.

The original Research-Plan-Implement framework compressed intent and code state into artifacts, but Horthy found reviewing both plans and code doubled human effort, creating anti-leverage.

He argues spec-driven development creates dual sources of truth that drift from code, making maintained specs often less useful than just using the codebase as reference.

Horthy credits the Department of Defense's 2018 DevSecOps factory essay for modernizing pre-AI software factories with automation stacks to accelerate government development cycles.

He traces the software factory concept to a 1968 NATO conference discussing systematized development steps akin to a factory floor, predating modern CI/CD and version control.

Horthy notes that reinforcement learning with verifiable rewards made Claude Code excel at tool use, but it optimized a single dimension, not long-term codebase maintainability.