April 27, 2026
Self-writing AI, self-starting drama
Tendril – a self-extending agent that builds and registers its own tools
AI writes its own tools—and the comments wrote the roast
TLDR: Tendril is an AI that codes and saves its own tools, getting smarter with each use. Commenters are split between fans who love the saved effort and skeptics calling it “fresh out of Claude Code,” with memes about the model “decompressing itself” and debates over real-world readiness.
Meet Tendril, the AI sidekick that doesn’t just fetch answers—it builds its own buttons. Ask it to grab news, it checks a little toolbox; if the tool’s missing, it codes one on the spot, saves it, and reuses it next time. In plain English: it learns your tasks and gets smarter every session without nagging you to approve it.
The crowd went full popcorn. Creator walmsles says they built it to stop agents “starting cold every session” and to solve the messy question of when tools should act on their own, not just how. One commenter nailed the meme-able vibe: this is “a mechanism for a model to de-compress itself.” Fans like gavinray cheered the savings—no more “burning tokens” doing the same chore twice—echoing their own “Saved Programs” at work. Others flexed with “we built similar” stories: nickstinemates touts integrated memory, repeatable artifacts, and keeping your current setup.
But the shade was strong: weitendorf joked “Get outta my swamp!” and called it “fresh out of Claude Code,” suggesting it’s still dev-only and not ready for civilians, dropping a rival attempt here. The drama? Is this a breakthrough in reusable AI skills or just another shiny agent memory? The community is split—half hyped by the “three tools to rule them all” swagger, half wary of an AI that auto-builds without asking. Either way, the memes and the rivalry are already shipping
Key Points
- •Tendril autonomously discovers, builds, registers, and reuses tools across sessions via a capability registry.
- •The agent core uses a minimal set of bootstrap tools (three in agent.ts) and accumulates capabilities over time.
- •A system prompt enforces behavior: search for tools first, build if missing, fix and retry on errors, and prefer live data via tools.
- •The implementation uses AWS Strands Agents SDK with an Amazon Bedrock model targeting Claude, and a Deno sandbox for code execution.
- •Architecture includes a Tauri (Rust) shell, ACP JSON-RPC over stdio, Node.js SEA and NDJSON, and a React frontend styled with TailwindCSS v4.