Alex Payne

@nicoalbanese10 This is amazing, aside from coding what are the first other problem spaces that come to your mind?

I've been iterating on this every day for several months and I'm so excited to open source it. While this is designed for coding, it can easily be adapted to other problem spaces. Fork it. Make it yours. vercel.com/templates/temp…

3 months ago I started building a coding agent that runs in the cloud. It's since written every line of code I've shipped, including itself. Today, I'm open sourcing it. Introducing Open Agents.

@i_am_brennan @davebcn87 haha! I think this is the prime kind of task that you could do at the end of a week before your reset to make sure you really get the maximum value. Perhaps have a scheduled task to improve all your skills.

@AlexPayne @davebcn87

GPT-5.4 is terrible at visual design. This article will try to convince you it's a user problem, but it has strong tendencies that it adores. Tonight's side project is to use @davebcn87's pi-autoresearch to develop a skill that makes GPT5.4 "Make {it} like Linear" Pray for my tokens.

@i_am_brennan @davebcn87 ok thats amazing!

In a violent display of burning tokens, pi edits the skill and calls `run_experiment()` which... 1. resets a sandbox folder & copies the new skills there 2. boots a headless codex there, and directs it to use it's skill to build {{Fixed Prompt}} 3. when codex returns, a script screenshots the page it made and another AI judges it against 6 or so reference screenshots and scores it against a rubric. 4. pi gets the average score from THAT ai (and shrunk images so it can see too)

@NeuralAA @robinebers well said

@robinebers But why Just be good faith and say yk thats the whole thing Anyways it don’t matter that much and they did good going back on it

ok lets fucking go unhinged cursor vs kimi thoughts

Have you cracked getting it to work with code review agents on PRs? Codex code review leaves an emoji reaction on the PR description and this changes from an 👀 to 👍 if there’s nothing detected. Not having to wait for these to run and then symphony fix, wait again would be great.

This works well for me: 1. Plan mode in Codex on xHigh 2. Copy markdown plan 3. Paste md plan into Linear issue and Symphony takes it all the way to human review (or auto merge of low risk)

@i_am_brennan Great article. Im a big fan of execplans. What are your usual targets for deslopification?

@odysseus0z @ryancarson I’ve been thinking about this too. Do you mean within symphony or something external to the symphony workflow?

@ryancarson (You can even add an agent to promote issues from backlogs to todo as well!

Holy shit, just had a feel-the-AGI moment. Symphony was running overnight, working on merging a bunch of PRs, and it encountered some extra issues in the codebase that were related to PR but not directly applicable. It created a backlog issue in Linear and documented it. 😳 What's even more amazing about that is now I can simply drag that issue into the "To Do" column in Linear, and Symphony will automatically pick it up, start working on it, and ship it.

@ryancarson @alex_frantic @linear I might have to do the same. Built my own in typescript but yet to test it out. Luckily the codex limits were reset multiple times so tokens weren’t fully wasted if I switch to elixir!

@AlexPayne @alex_frantic @linear copied the elixir ref almost exactly. much better than trying to do it in node

This is how I run 5 agents concurrently in a Code Factory to write/ship 100% of our code. It uses Symphony from @alex_frantic (oss) + Codex Mac app + @linear Took me 2-3 days to set up and now it’s *cranking* github.com/openai/symphony

@pvncher @TJHeinzer @RepoPrompt Wow!

@TJHeinzer @RepoPrompt I think it's the full context window - ~260k ish

It appears that the chatgpt team FINALLY fixed the low input issue with ChatGPT's 5.4 Pro. You can paste much larger prompts now and get full access to the model's intelligence. Going to be updating the @RepoPrompt context builder size token budget accordingly soon.

@pvncher @RepoPrompt Thanks Eric! I couldn't find any info on this in any of the docs/releases from today

@AlexPayne @RepoPrompt I'm inquiring. There appears to be a bug at the moment

Using GPT 5.4 Pro with @RepoPrompt is so easy now The agent takes care of building you a prompt, and you just have to hand it over to ChatGPT. You can even pass the result back for the agent to implement directly in the app!

@scaling01 Such a good model right now.

GPT-5.3-Codex Pricing: $1.75 Input $14.0 Output

@AnthropicAI If you are a good communicator nothing is stopping you being a power user.

New research: The AI Fluency Index. We tracked 11 behaviors across thousands of Claude.ai conversations—for example, how often people iterate and refine their work with Claude—to measure how well people collaborate with AI. Read more: anthropic.com/research/AI-fl…

@simonw it's so bad now. to the point where I have almost stopped going into comments as majority of replies seem to be AI generated.

AI-generated replies really are the scourge of Twitter these days. Anyone know if it's from packaged solutions being sold as a product or if it's people mainly rolling their own custom reply-bots?

@nicoalbanese10 @aisdk Sounds good thanks!

@AlexPayne @aisdk just local for now - we're overhauling telemetry and from that point will explore a similar adapter approach that would allow for different environments

New docs page to help agents get started with @aisdk ai-sdk.dev/docs/getting-s…

@nicoalbanese10 @aisdk That’s great, I was recently adding my own logs and then querying via the vercel cli to find out where latency was occurring so this should come in handy. Will it work in a vercel preview environment or only locally?

@AlexPayne @aisdk devtools are awesome! and we save outputs to a json file so super easy for model to query

@mattpocockuk I don’t understand what you mean by point number 5? The rest are all great. Top list!

I have been at 100% AI-contributed code for a few months now. Here are 9 ways it's changed my brain: 1. WAY more time thinking about integration testing 2. Friction via pre-commit hooks/CI/strong types is now super desirable 3. AI has no taste for UI, prototype extremely aggressively before committing to a PRD 4. AI has no taste for software architecture, be extremely explicit about the modules you want and think about their interfaces 5. Deep, grey-box modules with simple interfaces are the KING - let AI control what's in the box to decrease your cognitive load 6. Huge reliance on Effect.ts for dependency injection and strongly typed errors 7. Much more meta-programming, turning my skills into repeatable SOP's 8. First I thought MOAR DOCS = BETTER but 'doc rot' is real. Better to just let the AI generate JIT docs during plan mode. 9. Much higher cognitive load to keep up with the changes the LLM makes to the codebase