Insights on AI development, technology choices, and business outcomes
A tool claiming 98% less context use made us re-examine our own retrieval. Here's how context reduction actually works — and the one axis where their approach beats ours.
Meta entered the AI coding wars with Muse Code and Muse Spark 1.2. We went looking for the independently verified benchmarks, found none, and worked out what the harness-coupled headline score really means before touching our routing.
Every AI-generated landing page converges on the same skeleton. Evaluating an anti-slop skill exposed the fixed template hiding in our own pipeline — and the archetype fix that ended it.
Karpathy's "Claws" describes an always-on layer above LLM agents. We already run one — here's what the idea gets right, and the feature we refused to ship.
A viral engineering essay reminded us how easily an AI coding agent can "pass" a test by skipping it. We audited our own platform, found the gap, and fixed it.
Dify is a strong agentic workflow platform — but it solves a different problem than autonomous code generation. Here's how we evaluated both and how to choose.
Token-reduction tools promise 60–90% fewer tokens on AI coding, but cutting tokens isn't the same as saving money. Here's why signal per token — not raw token count — is the metric that actually protects quality.
Our AI development platform once pushed a live API key - despite having a secret scanner. The fail-open gap we found, and the fail-closed fix any AI pipeline can copy.
Multi-agent AI systems split work across a lead and sub-agents with isolated context. We build software this way — here's how super-agent harnesses really work and where the design choices matter.
