@ay_ushr Plan multiple much bigger tasks(6~10hr ones) with an interactive session where you're engaged, set them off and go enjoy life / scroll X for a couple hours and come back. Or focus on improving your harness / skills in the meantime
Live today on the Pepfessions podcast:
@huge_flood from Res Labs (@reslabs_ai) is joining us.
Securing the peptide supply chain: from the lab to your hands, verified.
→ A secure NFC tag on every product
→ Tap with your phone to check it
→ A record you can trust
Hosted by @troywhiteyyc
1 PM MT / 3 PM ET / 12 PM PT
If more and more services have stopped accepting tokens in URLs because they keep leaking into logs. Then why do we still accept RPC providers slapping your API key in the URL by default?
We will all be learning in coming weeks & months, that most human-made software:
▫️ Was really slow
▫️ Had many security vulnerabilities
▫️ Was under-specified and incorrect
AI will rewrite it all. Many times. It will make it fast, cost-efficient, and secure. It will think about failure modes we couldn't wrap our heads around.
It will prove, formally verify, load test, harden. It will turn memory-unsafe foundations into robust, optimized ones. If you ask, it'll thoroughly document it, teach you about it, make diagrams, animations, and slide decks.
In many ways, we are witnessing the true birth of software engineering.
i always disable all permission checks for my agents - i don't even do auto review. many people saw that in my videos and asked me how i dare do that, and that's actually a really good question
on a high level, there are 3 things that allowed me to do that -
1. a mindset shift - my machine is not mine
i now treat it as my employee's device. if i hired a human employee and bought a laptop for them, i would never tell them to ask me for approval each time they wanted to run a shell command on it
if you worked in IT for any big enough company, you would know that humans destroy their devices a lot more often than agents
the solution is not to introduce an approval process that constantly blocks my employees from doing their work. it's to be ready to wipe the whole thing any time, which leads to -
2. fully reproducible config
i showed this in my video and my public dotfiles repo - https://t.co/tWo1cOOBHZ
i use nix-darwin and home-manager to make my entire machine instantly reproducible. if anything goes wrong, i wipe the whole thing, clone my dotfiles, run rebuild, and a few minutes later everything's up and running again
3. isolate all secrets
production secrets and credentials should never live in plain text on an employee's machine. i always lock down all my secrets behind a gate that only i can authorize
i showed this in my latest video that my latest go-to solution for this is @AutomicVault - every time my agents need to access a secret, they go through me and i can decide whether i allow it or not based on the command they are trying to run that needs to secret
with this setup, my agents can work in a trusted by constrained manner, allowing them to move freely yet there's no way for them to create a disastrous outcome
And potentially dangerous instructions ie) "Act completely autonomously and stop asking me", work miles better when added as a text at the end of the prompt.
The agent seems ignores it when in agents.md
I'm pretty sure LLMs at least in frontier models, have a hierarchy when adhering to user instructions, and it goes
Last line of prompt > Other lines of prompt > Invoked skill > Passively invoked skill > Nested inferred skill > agents.md
Haven't tried a conflicting set of instructions to test this but thinking it would be a fun experiment
Is it just me or does @cmux text seem to disappear as the content goes up?
It's quite annoying when the post-execution summary is long and I need to reduce the character size to as-small-as-necessary to copy paste to text editor to read it.
Been a while since I've seen this by now.
cc @austinywang@lawrencecchen
Something I never expected to be spending a visible amount time doing: Cleaning up my 2tb mac for space.
I was spending about 5~10% of my time just cleaning up artifacts because when 5~6 sessions each with their own bundle of subagents are running wild, hundreds of gbs get filled in the tmp folder every few hours.
It got so bad I ordered a 4tb external drive and instead of a session cleaning up artifacts it's now an entire repo focused on cleaning my machine in real-time.
Curious how others handle this. FYI if you're not developing android / ios apps on at least one of the sessions you should be good
For long running tasks that run for 20+hrs,
I've yet to figure out how to keep the main agent to continuously run tasks with as many subagents as possible.
After <20hrs the agent usually ends up with max 2~3 subagents, and when I nudge it to parallelise more, it can spin up 5+ more, but ONLY after I tell it to.
Curious how others keep their main agents running on max parallel mode
Regularly going through the AGENTS.md files and hooks / skills setup of the top open source projects for agentic development(ie. opencode) I think is the best tip to make sure you're not falling too much behind.
Not sure if just me but Anthropic models / Opus is being super flaky. Getting a lot of random `Error: Anthropic stream stalled while waiting for the next event`.
@ClaudeDevs anyone else?
As a team we spent the past 8 years in crypto infrastructure, back when crypto was a gray market too and before the grifters and scammers moved in.
The lesson that stuck was that a gray market gets stronger when its legitimate players raise the bar from within.
RUO is at that same point now.
There's pressure building from the outside, from FDA attention to the growth of telehealth, much like when regulators and traditional finance started arriving in crypto.
The same principle holds here: when legitimate operators raise trust standards inside the market, the whole category gets stronger, and the good actors gain more of a say in how it grows and get to ride that growth long term.
If you see the gap between a COA and the vial in your hand as a real problem / painpoint, we'd like to work on solving it with you all.
Come join us!
Initially created this acc to follow and post about my experience with LLMs and agentic engineering because my main acc's TL was askewed to a specific industry
Since then this is what I've been working on with my fellow cofounders for the past few months
On October 14, it's time to say farewell to GPT-5.5 in ChatGPT, ChatGPT Work, and Codex across all plans.
If you use GPT-5.5 in Codex, switch to GPT-5.6 Sol or GPT-6 Astra.
Thanks for everything, 5.5 🫡