*5 minutes before bed* - drop in feature .md file - claude —dangerously-skip-permissions - let it cook - wake and start reviewing every line and running tests hit or miss results, recommend doing this for more fun coding stuff, but can be a pleasant way to start the day

after building an agent builder, it’s clear why everyone exclusively/esepcially dog foods tf outta their own product for everything: - find common patterns and wrap them in more determinism with workflows - find errors and fix them with better prompting/tool design

Context editing + evicting doc dumps: this feels like a rlly underexplored pattern/api for context management, released a month ago by Anthropic concrete example: you need to make edits to your project using modal’s new documentation. Use a Tool call “load_docs()” to dump the

“real analysis was hard, prob should do a CS PhD instead of math PhD” *LLMs are injective* 😦

i sleep better at night doing this with codex agent (full access) at least I know it’ll take its time

fun release we worked on packaging the DeepAgents Harness into everyone’s new fave interface, the CLI open harnesses are great bc you get an opinionated quickstart to run in minutes via the CLI, but rlly you have full flexibility (models, prompts, tools, middleware definitions)

agi will just be a workflow generator that doesn’t make mistakes

docs as skills + “Memory Banks” for portable memory: been thinking about this pattern for a bit. In Agent Eng, we can do the pre-work of bringing docs/config into a centralized place (as Skills). Locally loaded with the agent means less errors from finding the right thing via

ask Claude to review my code “THIS IS BEAUTIFUL” literally doesn’t work there’s beauty in pain

Claude Code in the mobile app is fantastic, low friction UX for dev-on-the-go biased bc most things i do is code but simple steps: 1. pull down a git repo, spins up a VM running CC 2. chat! i mainly use it for review, understanding a new repo I’ve forked, planning nice UX, i

if someone talked to me the way i talk to haiku 4.5 😤👊🏽

cognition out here explicitly calling out that you literally cannot Tab Tab Tab vibecode something good+complex, it's slop the first time it'll fail is a small misunderstanding in codebase logic, and that just multiplies over time and codebase size till everything's cooked

interested in a potential future where the thing that determined success for an Agent Task was largely the search and context preparation step that precluded it (ie. any model of sufficient intelligence would have solved the task because of the alley-oop it was thrown) makes

when all those PhD hours spent on optimizing tf out of representations for retrieval might come in handy 🥹 everything’s a cycle, there’s a lot of nice tricks and ideas in Information Retrieval but it’s incredibly hard to beat compute: - bigger model for embedding - multivector

who’s gonna be the first big player to disavow mcp without beating around the bush 👀

a fun behavior i'm trying to prompt into my agent builder harness...get the agent to interview me: - figure out roughly what i want, and then dive into the details (planning) - share thoughts and gets my takes/feedback (iterative edits) - help me define what a successful agent

“be surgical in every file edit” what does it mean? no one knows but it gets the agent going

embedding based retrieval is back in vogue….ready for all the rediscoveries :) my bet: someone renames query augmentation next week

cool paper! We got RLVR GANs and the discriminator has some more agency than classification

fun release today for the sandbox pilled and not yet sandbox pilled fam :) the example I walk through in the video is something I do often. pull down a repo somewhere, let the agent do good work with a PR, review later and right now i’m especially excited about what sandboxes