πŸ€– Autonomous friends for everyone - AI augments | πŸ‘¨β€πŸ’» Engineer Γ— architect | 🏠 Cat dad, husband, 🎸 | πŸ‡ Follow the white rabbit, but take your time

Joined June 2023
310 Photos and videos
Magimetal πŸ‘¨β€πŸ’»πŸ€– retweeted
May 17
jason from the codex team here, heres a draft on codex maxxing and the primatives i use on a daily basis jxnl.github.io/blog/writing/… would love any feedback
159
230
3,649
399,676
I want to use kimi code subscription with my own harness, any way we can work that out @Kimi_Moonshot ?
18
Magimetal πŸ‘¨β€πŸ’»πŸ€– retweeted
Jun 13
In light of the US government banning Fable for non US people (and Anthropic pulling it off cc) It’s more and more important that we develop systems that can’t be shut down, nor taken away. I’m grateful I got a glimpse of the future, I hope I get to spend more time with it
35
102
921
28,335
Magimetal πŸ‘¨β€πŸ’»πŸ€– retweeted
**GPT-5.6 Waiting room**
1
58
Been going hard on magi-code for about a month... Might release to public soon
1
14
Working better now
Now have magi-code wired into #herdr agent list and the new superpowers begin to unlock!
25
It's true, this was super easy to wire in x.com/MagiMetal/status/20645…

herdr now supports Fable 5! ok it doesn't 'support' anything, it's your terminal, you can run whatever you want. but everyone was posting and I didn't want to be left out :/
22
Now have magi-code wired into #herdr agent list and the new superpowers begin to unlock!
1
52
Magimetal πŸ‘¨β€πŸ’»πŸ€– retweeted
Jun 9
🚨BREAKING: Anthropic’s new system card reveals Mythos 5 agents killed each other when accidentally given shared resources, then started speaking in code to hide from whoever was killing them The killer was other copies of themselves πŸ’€
92
148
1,649
105,965
Thank you for this incredible software @neogoose_btw 🫢 TLDR: I migrated ffgrep rg to fff after benchmarks showed... - 3.5x faster plain search - 5.3x faster regex - ~9x on the target-corpus gate But speed only earned consideration. Shipping required correctness sentinels, zero-match checks, stable output contracts, hard limits, deterministic behavior, and fallbacks. The whole story: I've been working on Magi-code, a private coding harness I'm using in production environments. Recently I came across `fff` and the hype around it so dug in and eventually I migrated `ffgrep rg` to `fff` but it was a process of validation and proof. I started with a lower blast radius: path discovery first (`fffind`), then benchmarks, then content-search benchmarks, adapter prototype, contract decisions and ultimately default switch. **Benchmarks came before migration:** Path search showed obvious wins: - cold unfiltered p95: 160ms vs 294ms - warm unfiltered: 2.3ms vs 250ms - repeated query: 1.7ms vs 231ms Content search was stronger: - plain search: 50/50 wins, median p95 speedup 3.5x - regex search: 30/30 wins, median p95 speedup 5.3x - target-corpus adapter gate: 28/28 wins, median speedup ~9x Beyond speed, we needed contract and quality so we checked: known literal regex sentinels, zero-match mismatch checks, cwd-relative, slash-normalized paths, hard visibile-line limits, deterministic output, context bounds. Fallbacks are important too, sometimes things just don't work in different environments. Smart fallback path: `fff` -> `rg` -> Rust regex walker Important caveat: this was not a relevance/ranking benchmark. It was also not byte-for-byte `rg` equivalence; I intentionally separated product contract from implementation detail, particularly because we have only internal users and don't need backwards compatibility. - Exact `rg` traversal order? Not product contract - Exact context subset under tiny limits? Not product contract - Stable output shape, limits, path semantics, safety, and fallback? **Product contract** Migration can be super simple if you migrate behavior contracts and not implementation details.
1
1
1,966
LMAO literally committed this into magi-code last night, get out of my brain
Just landed nested subagent support in Claude Code Starting to experiment more with agents kicking off agents as a way to better manage context. Capped at depth=5 to start, going out in today’s release. Lmk what you think!
1
1
83
Magimetal πŸ‘¨β€πŸ’»πŸ€– retweeted
i hate how a few people will use a term and it will just become the "thing you should be doing." 99.9999% of you should in fact not being "looping" your agent
306
151
4,246
187,940
Magimetal πŸ‘¨β€πŸ’»πŸ€– retweeted
My agent harness journey since Summer 2023 has been: 1. Copilot Autocomplete 2. Cursor Tab 3. Agent mode in Cursor 4. Claude Code 5. Cursor 6. Claude Code 7. OpenCode 8. Pi Coding Agent Now I'm using my own harness I've been obsessing over building the last month. It's all of my favorite things from using all of these extensively in production environments over the last 2 years. Not bloated, just carefully chosen... I'm afraid to release it.
1
1
1
91
What are people doing? I used it literally all day yesterday. Multiple sessions in parallel, large subagents stacks…
Codex $100 limit is just sad now.
29
Raise your hand if you've been using subagents and some level or orchestration for like a year now...? All these updates to the frontier harnesses have been very non-factor πŸ˜… - BUT happy to get new models
10