Deep dives, practical guides and honest reflections on software engineering, distributed systems, AI and the journey of building in public.
Build a small retrieval evaluation set and measure context quality before spending days rewriting prompts or changing models.
AI & LLMs · Sep 27, 2026Design narrow, typed, permission-aware tools that make agent actions easier to validate, observe, retry, and audit.
Why agent infrastructure is moving beyond model calls toward managed runtimes, tools, state, sandboxes, and observable execution.
Long context is useful, but production AI systems win by managing repeated information, latency, cost, and execution location deliberately.
Tushar Sharma