You don't need an AI agent framework, or why frameworks are the new Juicero (konvu.com)

At a Konvupero lightning talk, the speaker made a pointed case: you don’t need an agent framework to build a useful AI agent — especially for v0. An “agent” is framed simply as an LLM plus instructions and a toolbelt; the runtime is a tight loop: prompt → model → maybe call a tool → ingest result → repeat. Practical hardening (keeping context trim, rate-limiting, validating outputs, handling retries, and adding telemetry) is essential, but none of it demands a new orchestration layer. The recommended pattern is boring and observable: a few typed tool functions, a simple loop, and metrics to drive iteration. The talk argues frameworks are often premature optimization — “the new Juicero” — because they rewrap primitives you already have (function calling, memory, tracing) and introduce a second DSL that leaks complexity. An illustrative “raclette reviewer” agent shows how compact, explicit code is faster to build, test, and evolve than state-graph frameworks. Frameworks earn their keep only when agent topologies and integrations are stable and repeatable across teams; until then, prefer plain code, native model APIs, and low-cost experimentation. In short: start small, measure failures, and add abstraction only when data proves it worthwhile.