Shoopy (@0xShoopy)

Yohei Nakajima, the creator of BabyAGI, just showed why he stopped building agents around the LLM. His AIE talk on ActiveGraph, in 5 timestamps: 1:55 – build around the log, not the model 3:24 – behaviors and policies gate what the agent can change 8:12 – his api key died at question 350, the run resumed itself at 353 11:11 – a loop that forks the agent and keeps a patch only if accuracy rises 15:49 – why long-running agents need an experiential world model The frame: make an immutable event log the agent, and replays, rollbacks and forks come for free. Different way to think about agents. 17 min well spent.

Anthropic's Lance Martin just walked through how they build agents that run for hours with no human in the loop. His talk on async, long-horizon agents in 5 timestamps: 3:19 – split the brain from the hands 6:20 – build agent vs verifier agent in a loop 8:18 – 20 iterations to clear an ML benchmark, unsupervised 13:19 – "dreaming" fixes memories the agent got wrong 16:08 – Claude Tag as an org-level harness, not a slackbot The point: past the 1-hour mark, architecture (not just model size) is what makes async agents work. Solid 25 min if you're building anything long-running.

Jeff Su, on why AI-designed slides all look the same: "Picking a template and prompting is exactly why your decks come out generic. prepare three files first, and every one comes out on-brand." In a ~14-minute walkthrough he shows the setup: → a DESIGN.md file (grabbed from a public GitHub repo, then cleaned up) as the foundation → turned into a reusable design system so the look is consistent, not re-decided each time → plus a template that gets better every time you give it feedback → the feedback carries forward – you're not re-explaining your taste every session → same workflow works for carousels and newsletters, not just slides The shift is the same one everything's circling: stop prompting blind, prepare the context once and reuse it.

A 4-agent pipeline ran loose for 11 days and burned $47K before anyone caught it. the fix is a pattern web teams have had since 2012, and almost no AI team uses it. Sachin Gupta (staff engineer, 15+ years on internet-scale backends), on why agents need feature flags: → most AI teams ship prompt, tool, model and autonomy changes to 100% of users on every deploy. no canary, no segment, no kill switch → that's why your "small" prompt tweak just broke 12% of users and you found out from a Discord screenshot → a boolean toggle isn't enough for agents. you need segment-targeted prompt variants, per-tool access flags, model-routing flags → the one that matters most: an autonomy-level flag – suggest vs auto-approve vs auto-execute – so you can dial an agent back without a redeploy → and a kill switch every agent system should have on day one, and almost none do Web teams stopped shipping to 100% on every deploy thirteen years ago. agent teams are about to relearn why, one runaway loop at a time.

Everyone's racing to remove themselves from the agent loop. the bottleneck just moved to the thing you can't automate away: understanding what the agent actually did. Geoffrey Litt (Notion) on staying in the loop without slowing down: → verifying the output is correct isn't enough. if you don't understand the work, you can't guide the next step, and the errors compound → he pulls this straight from education research: reading something and thinking you got it is the illusion. you only find the gaps when you're forced to explain it or use it → so he has the agent quiz him on its own changes, or writes the tricky part himself and lets the agent do the rest → his rule of thumb: the effort you save on typing should go into understanding, not into checking out entirely → the trap is that agents make it feel like you understand, right up until you have to touch the code yourself Going faster with agents doesn't come from understanding less. it comes from understanding more, in less time.