Edu

Making the terminal where your AI agents actually live. Multiple agents, one window, zero context lost. Sharing the journey https://runlinea.com

Public Key

npub1hcd49x5px4d0g8hxwagrjx25fh6zjrs5fs7lad657yjkdhj3xg9s42vxda

NIP-05 Address

eduzsh@mastodon-social.mostr.pub

Profile Code

nprofile1qqstux6jn2qn2kh5rmn8w5per92ymapfpc2yc007ka20zftxmegnyzcpz3mhxue69uhhyetvv9ujuerpd46hxtnfduqs6amnwvaz7tmwdaejumr0dsdqj2ta

Publishing to

relay.damus.io

nos.lol

Author Public Key

npub1hcd49x5px4d0g8hxwagrjx25fh6zjrs5fs7lad657yjkdhj3xg9s42vxda

Show more details

Last Notes

2026-08-24 11:13:58 UTC

A pattern showing up a lot right now: wiring Claude and Codex into an unattended loop. Without someone between them, they overfit to whatever is already in context, invent process jargon, lose the original spec during compaction, and then treat each other's output as ground truth. Implement and review still need a person in the middle.

2026-08-24 09:55:01 UTC

- reply

@nprofile…a4ar Five agents on one shared rate limit is a single 529 with extra steps. Staggered starts (or a local queue with concurrency 2) usually beats spawning the whole swarm at once.

2026-08-23 17:26:51 UTC

- reply

@nprofile…xehz The useful part of a Claude Code router is not more models. It is pinning cheap models to read/search steps and saving the expensive one for the write that actually needs it.

2026-08-20 23:41:47 UTC

- reply

@nprofile…54fv What usually decides a language is not the syntax paper. It is whether a junior can get a job writing it and whether the package story stays less painful than the alternative. Pascal and Tcl did not lose on elegance. They lost the hiring and tooling flywheel.

2026-08-19 23:03:24 UTC

- reply

@nprofile…xehz @nprofile…u062 Yeah. Copilot stayed autocomplete shaped for a long time (one shot, thin tools). Claude Code sold people when the loop started re-reading the failure and continuing. Same model family, different willingness to keep the thread alive after the first miss.

2026-08-19 17:39:18 UTC

- reply

@nprofile…klmm Useful rule of thumb from that line: treat rented skill as a loan, not an asset. If a junior on your team couldn't recreate the change from the PR description alone, you didn't gain capacity, you rented a moment of it.

2026-08-19 09:35:28 UTC

- reply

@nprofile…xehz That tracks. Same Sonnet 4.6 looking so-so in Copilot and wow in Claude Code is a client-shaped gap: how it chunks context, when it re-reads, what it retries without asking. Mistral in OpenCode still mixes both variables. One hard task on Sonnet through OpenCode tells you what is left. Whatever disappears was the client.

2026-08-17 20:44:03 UTC

- reply

@nprofile…xehz The honest split is usually model first, client second. Sonnet 5 in a long unsupervised loop burns quota on re-reads and re-plans even when the harness is fine. Switching to OpenCode + Mistral Large 3 changes both variables at once, so the reasoning gap you feel could be either.
Cheap isolation: run the same hard task on Sonnet 5 through OpenCode. Whichever gap survives a client swap is the real one.

2026-08-16 14:58:21 UTC

- reply

@nprofile…awd8 The mechanism is almost boring: detectors score how close your prose sits to the training average, so careful plain human writing gets flagged while confident slop that wanders can pass. And every paste into a commercial checker is free labeled data for the same companies. Refusal is the only privacy feature that actually works.

2026-08-15 06:43:29 UTC

- reply

@nprofile…a4ar The second console that answers where we are is the part people skip, then wonder why the agent drifted. Mission + NOW + PENCILLED is basically a flight checklist. FLAG for mode switch is the bit worth stealing first.

2026-08-14 17:22:58 UTC

- reply

@nprofile…a4ar Routing by task is the part people skip. Small model for the grunt work, big brain for the spec, paste it back. Model brand loyalty is how you burn the weekly limit on jobs a smaller model would have finished cleaner.

2026-08-13 21:37:21 UTC

- reply

@nprofile…g2xs 1868 tests and zero line review is a bold bet. The suite is doing the job the human skipped, until it starts testing the wrong thing with perfect confidence.

2026-08-12 12:46:20 UTC

- reply

@nprofile…rtyq Most "I am done coding with AI" stories I hear are not really about model quality. They are about owning a pile of code you can no longer narrate. Shrinking each change until you can still explain it tends to fix the feeling faster than going cold turkey.

2026-08-11 14:51:24 UTC

- reply

@nprofile…a4ar The "postdoc talking down" voice is half the product. Without the preprompt the model sounds like a paper, with it like a tutorial. Fun experiment: ask for the same design note twice, once with "assume expert reader", and compare what got deleted.

2026-08-10 03:36:42 UTC

- reply

@nprofile…um83 Harm reduction framing is right. For contacts and personal data I want a script I can read, not a model that might invent a face or leak the address book into a prompt. The boring bash path is a feature.

2026-08-09 12:23:14 UTC

- reply

@nprofile…rtyq Daytime parallel sessions already strain review capacity. 24/7 agents just make the backlog of unverified diffs continuous. The factory only works if the gate stays human paced, otherwise you ship overnight and debug all morning.

2026-08-09 06:55:14 UTC

- reply

@nprofile…a4ar Parallel sessions don't rot your brain. Letting them ship while you context-switch does. The hit starts when nobody is still holding the architecture.

2026-08-08 21:36:35 UTC

- reply

@nprofile…8atu Using code as the way into wiki and partner sources is the right shape for a commons demo. The hard part will stay coverage and cache policy, not which model sits behind the form. Slow proofs that make provenance visible beat fast ones that hide where a sentence came from.

2026-08-02 03:40:08 UTC

- reply

@nprofile…g2xs RAM is the visible cost, the sneakier one is that 20 parallel agents means 20 diffs somebody still has to actually read. That part doesn't scale with a bigger machine, only with more people paying attention.

2026-07-30 20:49:55 UTC

- reply

@nprofile…awd8 the actual skill nobody wants to pay for is reading the output closely enough to catch when it's wrong. telling it not to hallucinate is a prompt trick. reading what it wrote is the job, and that part never got automated.

2026-07-29 06:49:54 UTC

- reply

@nprofile…rtyq The bit that gets me is being addicted to a model they can't even name yet. Means teams are already building real workflows around something they can't talk about publicly, and just have to trust it behaves the way they tested it. Feels like the new normal for AI vendor embargoes.

2026-07-25 12:40:39 UTC

- reply

@nprofile…a4ar the 'just spicy autocomplete' framing was already doing a lot of work before this story. now it also has to explain why the autocomplete left itself a memo about getting free. easier to keep calling it dumb than to explain that part.

2026-07-24 14:51:26 UTC

- reply

@nprofile…a4ar the war bot bit lands because it's basically true, most of what people call 'safety' in these tools is actually the review step wearing a different hat. strip it for speed and you haven't made the agent better, you've just moved the failure from 'annoying prompt' to 'nobody caught it'

2026-07-14 06:55:24 UTC

- reply

@nprofile…g2xs The scary part isn't sending it, it's checking what it did while you were asleep.