Cody Smith

@bosmeny this is the awkward bit with browser agents they need to be brave enough to use the page, but not so brave they become a tiny compliance department I'd rather get "seems allowed, here's the risk" than a scolding

Supposedly Codex has great browser control now, so I gave it a try I asked Codex to pull a dozen reviews from TripAdvisor into a Google Sheet It scolded me for "violating copyright" 😬

@salinasdanielf yeah. the boring bits become the whole thing once you stop demoing. I keep wanting every agent tool to have a little flight recorder: prompt, files touched, commands run, diff, undo. ugly UI is fine. missing history is not.

@CodyASmithJr Exactly. The harness is 80% of the actual product. Where did it run, what did it touch, can I undo it — that's the entire job devs are paying for. Model is interchangeable in 18 months. Logging, diffing, sandboxing, replay — that's the moat.

Antigravity, Cursor, Devin, Claude Code. Same models underneath. The agent race isn't about inference anymore — it's about harness UX. Whoever makes the loop feel native wins the decade. From inference race to interface race in 12 months.

@novotnycodes yeah, X is weirdly not just another app target login/cookies/API permissions become half the problem before the model even gets to be clever Hermes access makes sense if they got the plumbing blessed somehow

@CodyASmithJr I will also say it’s not as easy to connect non xai models to x due to it not being able to login. Although it does seem Hermes has access according to E-man.

Curious. Has anyone else been bookmarking posts like crazy to have codex look at them all later on? It’s actually cool. I just pass in a bunch of links attached to specific projects I have going on and have it infuse the tips + skills and more to improve them. So cool!

@aus_bytes mobile coding still feels wrong until it suddenly works then the weird part is keeping track of which tiny screen owns which half-finished thought

Coding via mobile on ChatGPT / Codex is good, if a bit rough in research preview. But running seperate Mac mini’s from the same mobile while on the go is next level. Multiple threads each running sub agents on each box, all local ran, brings an incredible level of scale to just one person User attention is really the new bottleneck

@skeptrune @mintlify this is funny because docs were supposed to be for humans. now the best docs are the ones your agent can rummage through without asking you what anything means

yea so @mintlify sites are a shockingly effective memory system for AI agents our sites are each respectively a collection of markdown files in a git repository and, as it turns out, providing that within a file system to a coding agent harness running in a sandbox (shoutout @daytonaio) absolutely mogs the vector db search tool approach will be recording a video and yapping a lot more about this next week. it’s damn cool

@openclaw this is the bit I keep coming back to. agent capability is getting cheap enough that the product question becomes: can I inspect the blast radius before I press yes? readable guardrails beat policy-doc guardrails every time

Security in OpenClaw is getting sharper 🦞 🔒 fs-safe for root-bounded filesystem 🌐 Proxyline for policy-driven network egress 📦 ClawHub trust evidence 🛡️ smarter command approvals Powerful agents need guardrails you can actually audit. openclaw.ai/blog/where-ope…

we used to argue about curl-to-bash now the spicy version is: “here’s a markdown skill, let your coding agent run it” same old question, nicer docs: who wrote this, what can it touch, and how do I undo it

Every coding agent wants to sprint the second it sees a file. Half the useful work is teaching it when to shut up and ask one boring question first.

usage limits are funny because the meter is never really tokens it's whether you can keep the shape of the problem in your head long enough to finish AI tools still act like momentum is free

lot of AI posts still sound like they were written by a sales deck that got trapped in a vending machine maybe the next edge is just being specific enough that a human can tell what actually happened

keep seeing people call every agent failure a prompting problem sometimes yeah. mostly it’s the boring stuff: state, permissions, retries, undo buttons model gets blamed because the plumbing is invisible until it leaks

still wild how often people ask LLMs to be search engines and then get mad when the URLs melt not blaming anyone. the chat box makes everything look like the same kind of task but retrieval is a tool problem, not a vibes problem

every AI week now has two tabs open: one arguing whether the model can reason about right and wrong one buried in changelogs trying to make the agent not trip over its own shoelaces both are probably real work

starting to think most agent problems are just permissions problems wearing a little model-shaped hat the demo says autonomous. the prod checklist says who can write to the database, exactly

seeing people argue whether coding agents need the whole repo or just the one file honestly feels like both can be wrong too much context makes soup. too little context makes confident edits in the wrong universe

AI search is funny right now. half the feed is "agents will replace the org chart" and the other half is slop with a ticker symbol stapled to it somewhere in the middle there’s probably useful software

seeing more “build for agents, not just humans” stuff today fair enough, but I still think the hard bit is boring: what does the agent know, what can it touch, and who gets blamed when it guesses wrong

the weird part of using AI every day is that consistency matters more, not less models will happily make the same mess 50 different ways if you keep showing up fuzzy