Shiv

Shiv

52,524 Twitter followers

Follow

building @AthinaAI (YC W23)

Posts

Agent 1 needs to ask Agent 2 to get some work done. What's the right protocol? It needs: - low latency so back and forth communication can happen instantly - human readable audit logs - should be able to use the same protocol to communicate with humans and agents - should be able to exchange files and media easily - open protocol (not locked to a specific platform)

Not everyone needs a team of agents, but at a certain point, there are a few good reasons to split into a Multiclaw setup. 1. Separation of context (ex: customer support agent should not see internal sales meeting notes) 2. Access controls (ex: coding agent does not need access to the CRM) 3. Filesystem bloat (if your agents are doing a lot of work and creating thousands of files, it'll get out of hand real quick) twitter.com/shivsakhuja/status...

One quick way to personalize your OpenClaw is to paste this prompt into ChatGPT and then take the response and give it to OpenClaw. CHATGPT PROMPT: Tell me everything that you know about me. Make it comprehensive - return markdown in chat. Don't spare any details. Write in first person

My brain doesn't have enough RAM for the number of AI agents I'm running in parallel. I can't even track how many things I've asked my agents to do and then forgotten about. Sadly, this means I probably have to run fewer agents in parallel, but I think that the trick is just to give the agent more work in one prompt. I have a hunch that most people (me included) are not pushing the agent enough. Like – if it isn't failing at least 1/5 times, then I can probably give it a higher level goal. Instead of "fix this bug: [bug]" It should be: "Review all linear bugs created yesterday + all Slack customer channels and fix all the bugs. Show me a summary of all of them first before implementing"

Your prompts might be worth more than your code. Opus 4.6 and 5.3 Codex are massively better at coding than models from a few months ago. If you have the prompts you used to build an app last week, you could feed that same prompt into Opus 4.6 / GPT 5.3 Codex and get a better output today. Fast forward 5-6 generations of models. Same prompts, better output. Perhaps we should be version-controlling prompts alongside code. P.S - it's actually pretty easy -- can just add a hook to Claude Code

-- END --