I am desperate for feedback on my homebrewed harness, "Plurnk."
I've been dogfooding it on my RTX5070Ti with only 16GB VRAM as my daily driver for weeks now and it's genuinely better than everything else available (I'm biased).
- Curation, not Compaction
The model's fully responsible (bitter lesson) for deterministically curating its own context.
- ANTLR Grammar
The dozen verbs for the harness are optimized for model pretraining, looking like markdown while being fully integrated with an EBNF grammar that plugs into an AST for superior workflow and recovery.
- Universal Resources
Everything's addressable by the model through a pseudo-URI interface, including the log entries.
- Omnipatterns
Everything in the repo goes through treesitters which build a complete graph that the model can search with glob, regex, jsonpath, xpath, graph, and sqlite fulltext.
- Standards
MCP2 tools, AG-UI client interface, A2A "agent to agent", plugs into everything through OAI spec + models.dev
- Interfaces
CLI, TUI, and a Neovim plugin, easy to extend, very pluggable
- Lean Agent
Service/Client architecture with lean TUI allows dozens of agents to run on weak hardware. 200MB footprint.
- Lean Context
Achieves all of the "batteries includes" features of a fat harness while the sysprompt is only slightly larger than pi agent's.
It started as a proof of concept for my bespoke opinions on what "bitter lesson" actually means, and it's gone so well that I believe others trying to build agentic workflows, especially with local and humble constraints, could benefit from the proven concepts.
Comments URL: [https://news.ycombinator.com/item?id=49720466](https://news.ycombinator.com/item?id=49720466)
Points: 1