Skip to main content

Agentic Frameworks

Which framework fits the job, and when should you skip frameworks entirely?

The Spine​

  • UI from prompts: v0 — move quickly from a wireframe to a component.
  • Full-repository coding: Claude Code — work autonomously with repository context.
  • Verified capability adoption: Codex — choose a surface, bound authority, and test independent engineering capability.
  • Context-aware editing: Cursor — keep context and changes inside the editor.
  • Crypto agents: ARC RIG — use a Rust-native, Solana-oriented framework.
  • Social AI agents: ElizaOS — work across a multi-platform plugin ecosystem.
  • No framework: Direct API — retain direct control with fewer abstractions.

Choose by the work, evidence, and responsible reviewer. A larger framework is useful only when its coordination and control repay their cost.

Zoom Out​

  • Agents — decide when the job needs an agent rather than a bounded skill.
  • Autonomous Agents — understand framework concerns for production agent systems.
  • Prompts — improve the instructions every framework must interpret.

Changes my mind: a direct, lower-complexity route repeatedly fails the same bounded job while a framework closes the gap without hiding authority or proof.

Context​

  • Tight Five — connect the chosen framework to a question, principle, protocol, standard, and platform.
  • AI Coding — put the selection inside a repeatable engineering workflow.
  • Agent-Operable Codebases — test whether the codebase can support bounded agent work.

Questions​

Next question: at what complexity does framework coordination save more energy than it costs?

  • Which framework assumption breaks first beyond prototyping?
  • When does the abstraction between the engineer and model become the bottleneck?