@dvyio I’ve gotten so addicted to subagents to save the main agents context window.
but i use a Goal Ledger SKILL by @KingBootoshi so i could probs get by with less subagents for long runners
long running tasks in codex are broken
love the extra resets but my lead scraping workflow (max of 3 subagents) is unusable.
anybody else getting this “run away application memory” w/ Codex?
the Hermes + Codex upgrades i made this week:
> Project Memory - tells the agent how this project works:
- recent work, decisions, rules, taste, defaults, etc...
> @honchodotdev remembers user + agent prefs across sessions
> codebase-memory-mcp gives the agent a real map of the repo.
think about it like this:
the "workbench" around the LLM is where agents go from cool chatbot with tool calling to full gigachad agent.
https://t.co/8WlK0wXD0N
the Hermes + Codex upgrades i made this week:
> Project Memory - tells the agent how this project works:
- recent work, decisions, rules, taste, defaults, etc...
> @honchodotdev remembers user + agent prefs across sessions
> codebase-memory-mcp gives the agent a real map of the repo.
think about it like this:
the "workbench" around the LLM is where agents go from cool chatbot with tool calling to full gigachad agent.
https://t.co/8WlK0wXD0N
Hermes Agent v0.18.0 is here. The Judgement Release.
This release includes a lot of powerful new features:
- Mixture of Agents as a first class virtual model, to take you beyond the frontier
- /learn to teach your agent anything from open ended sources
- Journey to visualize how your agent learned
- Gemini Vertex is now a supported provider
- Fable 5, Sonnet 5, and Fugu are now available
- Projects support in the desktop app to focus in and build
and a ton more, including our team clearing ALL P0 and P1 issues for the first time since our project took off!
Check it out:
codebase-memory-mcp makes Codex less drunk in big repos.
it indexes the codebase into a graph, so the agent can ask relationship questions instead of grep-goblining through files.
what calls this?
what route owns it?
what breaks if we touch it?
sick as fuck.
agent made UI needs a taste check.
i updated the compicker SKILL so you can make agents build 3-7 real UI options, wire em into the app, and test each one in your browser.
> generate options with agents.
> choose with taste.
> lock the winner.
> drop the picker.
SKILL ->
my favorite UI trick in Codex right now:
run $image-gen before writing components.
make 3 ugly-but-directional screens, pick the vibe, then tell Codex to build toward that.
half of bad AI UI comes from giving the agent no taste target and praying.