so models are coming prebuilt with memory capabilities?
if that’s true- then we’re getting to a point where a lot of the harness work will be solved from the foundation
i've rebuilt our agent harness countless times and each time i come back to studying how the best ones are built
harness design is more important than every before. as models get better, the agent harness needs less scaffolding and customization
ultimately - they all do the same thing, one core loop that every front door ends up calling. the edges are what makes them unique
this covers:
- hermes agent by @NousResearch
- @opencode
- @pi
- goose by @TheBlockCo / @AgenticAIFdn
- deepseek harness by @deepseek_ai
- codex by @OpenAI
any agent can benefit from these standards without reinventing the wheel:
- pi and dsh have small core and set up everything as a plugin. whereas codex and hermes build out everything has features
- hermes builds the system prompt once and reuses it so the provider cache survives the turn
- compaction is designed with the same basic strategy, summarizing older history and tune it differently
i go down rabbitholes on pretty often and i usually find myself in a pickle not understanding concepts or terms
so i made this mac app to make it easier without opening up chat/claude
pickle assists in the understanding -
- simplify/expand articles and passages
- explain a term with context
- can use the page and links around it
- runs locally and online
- creates a knowledge base to track progress
i rarely use the cli nowadays but this may change my mind
you can now watch videos while claude is running
connect your youtube then watch videos or shorts
i built this out watching intersellar on the side
Claude Code just shipped mods...
I made one that connects you to a multiplayer Doom server while Claude is busy.
Every other player in the match is a person waiting for their Claude to finish.