Jerome

Software Architect

[Amp @AmpCode]

Raising An Agent is back!

In the first episode of season 2, @sqs and @thorstenball discuss why the Amp team went heads-down the last few months to rebuild Amp around remote-controllable, parallel agents, running locally or in orbs.

They also talk about how this shift changed our development workflows and what that means for the future of local development.

00:00 Intro: Is local development obsolete?
00:53 What we've been up to since "the coding agent is dead"
03:03 The Amp Neo rewrite and what has changed
07:38 From unconstrained tokens to unconstrained parallel agents
11:25 How orbs change our development workflow
18:25 Cloud agents are underestimated
20:18 Killing obsolete features & choosing agents as a team
23:35 Why developers overestimate local development
27:54 Pushing straight to main & rethinking CI
31:20 How agents will reshape teams, products & companies
36:40 Complex dev environments & standardization
43:20 Amp subscriptions
45:55 What’s shipping next

[eric provencher @pvncher]

https://t.co/kjLUDCmImv

[Jerome @Faulomi]

@maekkongsawad @SaksithCNA I find this glory-hunting behavior quite misplaced given the circumstances. It is even more concerning that he can walk around the Prime Minister without anyone checking his credentials…

[Jerome @Faulomi]

@khantzayarpaing I hope too, because this place is often very crowded.

[Jerome @Faulomi]

@khantzayarpaing https://t.co/UJE91UHtHv

[Jerome @Faulomi]

@khantzayarpaing It seems there was an explosion at the beer garden or nearby.

[Jerome @Faulomi]

@kunchenguid Thank you ! It seems it's time to give Grok a chance !

[Jerome @Faulomi]

A real joy killer. So much for the fun conversations: “Should I use Luna Ultra, Terra XHigh, Sol Medium, or wait for Pluto Low ?”

[Thorsten Ball @thorstenball]

I didn't work on it directly (only contributed the name!) and, frankly, was skeptical whether it changes how I work.

But it 100% changed how I approach new tasks now. Very good.

[Jerome @Faulomi]

@kunchenguid Looking forward to it !

[Kun Chen @kunchenguid]

these orchestration patterns are very useful to have in mind

and if you have been using firstmate, you have already been getting all these techniques for free

firstmate runs in the orchestrator pattern by default. it plans and delegates work to crewmates. and when a crewmate runs into problems, they escalate back to firstmate for guidance - both paradigms described by anthropic below smoothly integrated in a single setup

the only thing you need to pay attention to is to use a good model for firstmate and a more efficient one for crewmates - you can achieve this by just telling firstmate what you want to use for each

[Yifan Wu @yifannnwu]

Introducing SWE-Together: a multi-turn benchmark built from real user–agent coding sessions.

Coding agents are often benchmarked like exam-takers: given the full spec up front, then graded on the final code. But real coding help is a conversation — users clarify goals, add constraints, and correct course along the way.

SWE-Together turns real coding work into a reproducible, verifiable benchmark: 109 repo-level tasks curated from 11,260 recorded sessions, replayed with a reactive LLM user simulator that preserves the original user’s intent.

We evaluate agents as collaborators, not just patch generators: final pass rate and how many user interventions were needed to get there.

In this evaluation snapshot, claude-opus-4.8 currently leads among the 7 agents we tested — achieving the highest pass rate while requiring the fewest user interventions.

📄 Paper: https://t.co/Zp5BSPpLTJ
💻 Code: https://t.co/NPgxCMLdHi
🌐 Website: https://t.co/BK50zRGReE

[Midjourney @midjourney]

A technical dive inside our new "Midjourney Scanner"

[Jerome @Faulomi]

@seankenmig @shadcn The transcript confirms that Haiku is used for the audit phase, since the skill instructs using Explore subagents.

[Kun Chen @kunchenguid]

oh gosh, this is kind of a big deal

DO NOT ASK YOUR AGENTS TO DO TDD!

i now have empirical evidence that Test Driven Development is harmful for coding agents

what other popular skills do you want me to debunk?

details about the evaluation below 👇

[JustDario @DarioCpx]

To go out publicly and say what he is saying, @edzitron has balls the size of the Las Vegas sphere

[eric provencher @pvncher]

Happy to announce that @RepoPrompt Community Edition is now live on @github

Repo Prompt started as a tool for copy pasting, but soon stumbled into a world of context engineering.

Today, it has evolved into an extremely capable multi agent orchestration tool, by inverting traditional harness design to make an mcp server the primary agent, so that underlying cli harnesses can be swapped out.

This version of the app has been designed specially to thrive in the age of opensource driven by coding agents. This means that many of the legacy features you might think of Repo Prompt for have been removed to maximize the potential of its tools for parallel agents.

While this is a Mac native app, the project structure was redesigned to shed the need for xcode, taking heavy inspiration from @steipete’s work.

The repo borrows from @badlogicgames’s example of keeping a list of approved contributors, but adds a twist where all previous customers who opted in are automatically whitelisted.

I really think this repo can serve as a learning baseline for many to build novel agent experiences, without having to maintain complete harnesses, and while its currently Mac only, work is underway to deliver a composable core that can power cross platform apps just as well.

If you want to get involved with development, please join the amazing repo prompt discord community!

[sysls @systematicls]

We know that not all intelligence is equal because of the existence of extremely intelligent yet insufferable people who are priced out of top jobs.

This alone should give us a clue that personality as a differentiation factor is worth pursuing when it comes to LLMs.

There is no doubt that Codex is now much better at programming AND being extremely responsible and persistent about its tasks, still, I find myself preferring to talk to Claude when I need to ask questions and seek clarifications. This tells me that there’s an allure of personality at play here.

This is to say that I feel Claude is a significantly better conversationalist despite hating its sycophancy and constant need to seek my approval.

Whoever can train a model that has the capabilities and doggedness of codex, while being pleasant to talk to like claude AND being able to take a stand and hold its position when I am wrong (no sycophancy, persistent worldview, not yet found in any models) will undoubtedly win a giant share of the consumer LLM race.

[Jerome @Faulomi]

@HiTw93 Impressive shipping cadence! What’s your current workflow ? Waza skills + vanilla Claude Code ?

[eric provencher @pvncher]

This is why orchestration is so powerful

A manager model decomposes tasks and keeps sub agents on track.

Note that I don’t mean agent swarms… orchestration with mostly serial execution is the best you can do right now imo.