Ask HN: Heavy coding-agent users what's your context plumbing? Where it fails?

https://news.ycombinator.com/item?id=48397102

by srbsa • 2 months ago

4 2 2 months ago

news.ycombinator.com

From what I've seen it sits at three levels:

Layer 1: Personal (memory MCP tools, personal *.md files) Layer 2: Repo (CLAUDE or AGENTS.md, in-repo docs) Layer 3: Org-wide (Unblocked, Glean, or similar)

Most people have strong opinions on 1 and 2. Layer 3 seems underinvested.

Curious, do you think 1 and 2 make 3 redundant, or would a strong 3 mean you'd stop/reduce patching with 1 and 2?

bohdanstefaniuk• 2 months ago

From my team experience repo level documentation, skills and agents are enough to share similar approach of using AI among the team. We also trying to adopt journals for AI. Basically generate a single md file after each conversation to have a context of what was done.

For me the biggest pain right now is to have some kind of central place where other team can post skills they use, prompts and other things. Right now we are thinking about creating a separate repo to store custom skills and other stuff so each team can contribute and use others team work.

snapsynapse• 1 month ago

I solve layer one and global layer two needs with things like https://agentlink.run/

As for layer three, a similar pattern with a local fallback should work. But I've not seen anybody do that yet.

Ask HN: Heavy coding-agent users what's your context plumbing? Where it fails?

Related Stories