Chris Raethke

@mattshumer_ I've been working on the early stage of mostly the same thing (for my own engineering needs first), but broadly applicable teamcadence.ai/blog/context-a…

Agents that natively self-orchestrate, managing their own context, tools, and sub-agents, are the next big unlock in LLM performance. Right now, a skilled engineer building an optimized harness, with thoughtful data flow, separation of concerns, sub-agent management, etc., can make dramatic improvements over baseline for specific tasks. If a model could do this itself, that’d be a major step forward. You give it an objective and a set of tools, and it figures out the optimal way to orchestrate itself to do the task. For example, I’m building a very primitive AI scientist that I’ll open-source soon. Most of the work isn’t in the prompt, it’s in the harness… what the orchestrator sees, what sub‑agents see, what gets shared between them and when, where we summarize vs. pass raw data, and which tools each agent controls. Doing this allows me to dramatically improve what the model can do on its own. If a model can effectively design its own harness for a given problem, it’d be a huge step forward. My bet: self-orchestrating models… ones that manage their own context, tools, and sub-agents, will move the frontier almost as much as the jump from chatbot → reasoning did. Maybe more.

Your coding agent accumulates patterns you never notice. Skillable mines your session transcripts to find repeating workflows and turn them into skills. Works with Claude Code, Codex, OpenCode + 9 more. github.com/codesoda/skill…

@zarazhangrui I actually built a skill, which analyzes your agent usage patterns and then identifies any patterns that would suit custom skills for you. github.com/codesoda/skill…

How I build skills: 1. Brute-force a task with Claude Code 2. Iterate until it meets my criteria 3. Get Claude to read this folder, where I pre-downloaded blogs & docs from Anthropic team, and turn what we just did into a skill 4. Eval the skill using skill-creator plugin

I love that I can just follow along as it builds it out. github.com/codesoda/agent…

I worked with GPT 5.4 to brainstorm, then codex and I wrote a PRD (see tasks). That was then strategically broken up into detailed specs using my auto-ralph skill, which I reviewed meticulously. Now auto-ralph is [planning, buidling, reviewing]+ through each spec to build it out.

Currently ralping a new thing AI first bookmarking - What's interesting is not so much in what it is, but how it's coming into fruition. github.com/codesoda/agent…

@camhahu In theory, but I feel the value (and current need) is more on the vibe-coded end of the AI maturity spectrum

@codesoda time to rename to agentic-engineering-audit? 🤔

not having being an engineer is an advantage... but you need these 5 things that everyone who's a great engineer has

@camhahu A vibe code audit come up with this scope. My ralph PRD skill broke it down into specs. Now ralph is building it all planning, implementing, reviewing and pushing step by step to github.com/codesoda/vibe-… This will be better code and tests than I could dream of being able to write

what if humans only reviewed changes to the CI, not the code? we decide what decides what passes feels scary but directionally correct

@camhahu Even better with proper embeddings, do you have a log you can DM? I can have a look.

@codesoda looks like mine fell back to just bm25 and it's not clear why - but even with that, these suggestions are really good

every great engineer I know who isn't token-maxing right now dislikes AI because it make you worse before you figure out how to use it right

@camhahu github.com/codesoda/vibe-… Try this it's killer

@codesoda what have you been doing for audits? particularly on stuff where you're doing little/no manual review we've looked at some stuff like cyclomatic complexity and module boundaries but no silver bullets

@Al_Grigor If you are using terraform then make sure you have a thorough way to test your setup, and make sure it's documented in your AGENTS.md. I'm biased but the best way to do this is with my terra-agent skill from the x-agents project. github.com/codesoda/x-age…

Claude Code wiped our production database with a Terraform command. It took down the DataTalksClub course platform and 2.5 years of submissions: homework, projects, and leaderboards. Automated snapshots were gone too. In the newsletter, I wrote the full timeline + what I changed so this doesn't happen again. If you use Terraform (or let agents touch infra), this is a good story for you to read. alexeyondata.substack.com/p/how-i-droppe…

@camhahu Or running vibe code audits over night, so I can then work on improvements the next day.

@camhahu I get to the end of the day and "oh what! I have tokens left before I'm rate limited, time to fire off some ralph loops on some smaller lower priority work items"

@camhahu Love this, andressen had a similar comment on his recent interview. Engineers do so much more than punching on a keyboard.

think of something you want to build, the fastest way is with an LLM, but the LLM still needs you therefore: coding is solved, software isn't

@ryankoonce @garrytan ..then have codex check the plan, then give codex's feedback back to claude, repeat.

@garrytan Plan mode. Then have Claude.ai check the plan.

Here's my process: Spot a problem while using the product (or have a feature idea) Describe the architecture to Claude in enough detail that the first draft is 80% right Review the output, catch the subtle bugs (ordering of side effects, race conditions, security issues) Iterate 2-4 times within the same PR (your PRs average 3-5 sub-commits, each fixing something the previous round missed) Ship it — version bump, changelog, merge Do you guys follow this or do you do something else? I'm curious

You know what would be nice, different permission profiles inside a @claudeai project, e.g. I'm in dev profile, so that has a bunch of skills, tools, etc., enabled. I'm in the AWS/DevOps profile, so I enable all the default things for that. But in general, it's locked down.

@gdb I'd love a way to skip the sandbox when launching though, I'm often doing a bunch of ops type things which means actually making mods to the local machine.

codex app has gotten me to switch from emacs / terminal for the first time

@zarazhangrui There is certainly an echo chamber going on

Seems to me that every new AI product is trying to grab the attention of a very small group of elite, AI-savvy users who are completely bombarded and tired of their marketing tactics While at the same time normal people who are outside the AI circle are completely oblivious to the latest developments in AI & only using it as a google replacement

@tanmaye_bhatia @nicbstme 👋

@nicbstme Interesting! I think there needs to be step function to have dynamic and reliable UIs in agent windows. They’re not good at this right now. If you want to action on data, buttons and clicks are more reliable that md files. I wonder if any companies are building here