prime-agent vs whip
Self-improving coding/research agent around a Recursive Language Model: persistent IPython as the core tool, programmatic subagents, durable harness state it refines via evidence-backed /refine. — versus — WhipCode: open-source coding agent in Go built on a recursive language-model loop; agents delegate to sub-agents through short programs, with daemon-owned sessions across desktop, TUI and web.
Both build the coding agent around a Recursive Language Model: prime-agent's core tool is a persistent IPython with programmatic subagents, WhipCode's loop writes Starlark/JS programs and delegates through a shared daemon.
| prime-agent | whip | |
|---|---|---|
| Stars | 21k | 1.1k |
| Forks | 2.3k | 153 |
| Language | TypeScript | Go |
| License | MIT | Apache-2.0 |
| Last activity | today | today |
| Topics | coding, agents | coding, agents |
| Curated connections | 7 | 3 |
prime-agent — the curator's take
The two abstractions are genuinely different: context as variables in a persistent REPL (not a transcript), and a harness that rewrites its own supplemental state with rollback — the closest thing to a shipping continual-learning agent. Daemon sessions and agent-to-agent messaging make it a long-horizon tool, not a chat CLI. When NOT: it's a full worldview — you adopt the RLM way or fight it; young codebase moving fast, and self-improvement means your harness drifts from everyone else's.
whip — the curator's take
Interesting when your tasks outgrow one context window: instead of a growing transcript, an RLM loop writes small Starlark or JavaScript programs to slice history, files and large tool outputs, and delegates recursively to its own sub-agents, while a daemon keeps sessions alive after you close the window. Built with open-source models in mind, with provider routing, MCP and skills. Its benchmark (20 of 30 on a mixed Terminal-Bench and repo set) is honestly caveated as not a matched comparison. The churn is real: a v1 alpha CLI track, an Apple-Silicon-only desktop beta, and a clean-project reset that is not an upgrade path. Stay on Claude Code or Codex for a stable daily driver; try this when long-horizon delegation is the actual problem.