Andrew Mackross

I feel startups are the same. When you’re small the problems are hard, when you’re big the problems are hard. I think it’s just the nature of problem solving the limiting factor. Now AI is taking more of the easy things — we’re left just with just the hard problems it can’t solve (or getting it to solve is not worth ROI). Thus the problems are harder and the competition can solve all the easy things too now. More than ever differentiation on hard problems is required.

think back to projects you've worked on in the past it's hard not to imagine they'd have been completed way faster now that we have ai but everything still feels as slow and as difficult as ever

@Schappi Oh super nice!! Yeah that’s hard lol.

@Schappi What use case?

If you’re willing to burn ~30% of a cpu core on a custom semantic VAD and hack through the bugs in gpt-realtime2 you can build something that feels much more responsive and natural than OpenAI semantic VAD and as a bonus it keeps your computers toasty. Stack: WebRTC carries full-duplex opus audio from local server that also connects with WebSocket to GPT Realtime2 with OpenAI VAD disabled. Local server decodes the different PCM/sample-rate paths for all the different detectors+openai and also encodes for browser playback of assistant voice. Server runs local Silero VAD (ML), assistant echo-aware barge-in gating with non word utterance detection and auto continuation. It uses a tuned multi-checkpoint Smart Turn threshold curve (smart turn is a ML model for end of turn detection, but running it 7 times at different times is much better). Server playhead telemetry drives deterministic interrupt, truncate, cancel, and context-repair logic, and works around Realtime API bugs and edge cases.

@zebassembly I tried this but found that it was still worth having apply_patch on edit with gpt5.5 using lark. Way more reliable as it doesn’t have to json escape everything.

what if your agent was entirely codemode

@threepointone You’re cranking

Scheduled Tasks for Project Think github.com/cloudflare/age… - use cron patterns or a DSL - run a prompt (or regular code, coming soon) - ... that's it. simple.

@rasmus1610 Yeah nice, I’m doing something very similar to create package expert agents

This feels like a automatically managed AGENTS .md file. I think this can be a great way for LLMs to continuously learn while interacting with context

Little trick for harness devs that has worked nicely for me, when i need to inject important “pushed” information mid conversation, i create a fake tool call request and result for the pull version of that thing. Like get_modified_files, or get_chat_notifications. Agents seem to trust tools more than injected system messages.

@tunguz lol timely, I just posted how if you give up about 30% of a core to custom semantic VAD + encoding/decoding/resampling/detecting you can make your realtime voice assistant feel a lot more realistic.

Here is one big reason why this matters. Time spent on non-LLM inference tasks is only going to increase. However, tools that these AI system use are *very* inefficient and have been built from the ground up for CPU and human use. There is a huge untapped opportunity there to significantly improve those processes with AI agents in mind from the ground up.

@antigravity teamwork-preview hitting rate limits every second 30s after first use (brand new ultra customer)

@tunguz Is it the consulting company or the product that you did?

I did the thing.

@dok2001 CF has so much good stuff but seriously bad at telling story / price bundling

What are we missing?

@threepointone Somewhat similar I’ve been working on a join where the main agent decides it’s got enough info from subs and writes a summary that ends up replacing where the fork started off.

tl;dr - subagent behaviour working on adding multi chat and subagents to the agents starter (yay!) and I have a curious product direction/question. our subagents can be full fledged chats themselves. which means they could not only be async while they work on their thing and you continue, but you could continue "talking" with them after they've "returned" a result. so what should the default behaviour in the starter be? - readonly, no input. this is what most (all?) products/devtools like this do atm - have chat, but it's only followups, doesn't affect the main chat - add a "send back/summarize to main chat" this feels powerful and underexplored I'll probably ship option 1 for now, but there's something here... anyway, multichat/subagents in starter template coming this week

@r0ck3t23 Yeah but do I want an AI dog therapist for pennies on the dollar or a human one who costs $20/hr, is only available in work hours, and can only serve one customer at a time. Even if there are new jobs, they are not going to us meat bags. Clueless logic.

Jeff Bezos asked a room to imagine going back a hundred years. When almost everyone was a farmer. And telling those farmers that in 2018 there’d be a job called “massage therapist.” Bezos: “They would not have believed you.” Then a friend took it further. Bezos: “Forget massage therapist, there are dog psychiatrists.” He looked it up. Bezos: “Sure enough, you can easily hire a psychiatrist for your dog.” The room laughed. The point under the laughter wasn’t funny at all. Every time a major technology shift hits, we do the exact same thing. We count the jobs it will destroy. We never count the ones it will create. Because we can’t. They don’t have names yet. The fear is always specific. AI will replace accountants. AI will replace radiologists. AI will replace drivers. The fear has job titles and timelines and projections. The opportunity has none of those things. Because you can’t name what doesn’t exist yet. A farmer in 1920 could understand losing his job to a tractor. He could not understand gaining a career as a social media strategist. Not because he lacked intelligence. Because the entire chain of inventions between his world and that job hadn’t been built yet. Radio. Television. The internet. Smartphones. Social platforms. Creator economies. Every single link in that chain had to exist before “social media strategist” could even be a sentence. That’s where we are with AI right now. Everyone is staring at the tractor. Nobody can see the thing seven inventions away that doesn’t have a name yet. The fear is loud because it fits inside language we already have. The opportunity is silent because it doesn’t. Every technological revolution in history created more jobs than it destroyed. Every single one. Not because anyone planned it. Because human needs expand faster than machines can fill them. We didn’t need massage therapists when we were breaking our backs on farms. We needed them after machines freed our backs and stress replaced labor. The demand didn’t disappear. It migrated somewhere no one was looking. That is exactly what’s happening right now. The jobs AI creates won’t make sense to us yet. They’ll sound as absurd as “dog psychiatrist” would’ve sounded to a farmer in 1920. Until someone is running a $200 hourly practice with a six-month waitlist. The entire conversation right now is about what we’re about to lose. Nobody is talking about what we’re about to gain. Because the gains don’t have vocabulary yet. A hundred years from now, someone will stand on a stage and describe the jobs we couldn’t imagine today. And the audience will laugh. The same way we just did.

@Cloudflare I’m long you guys (big % of my portfolio), but you need to sort out your go to market.

@tunguz What’s with everyone throwing out basic reasoning… you can’t use history as a guide when the underlying assumptions are completely fucking different.

It is different.

@mitsuhiko Delegate pattern and interfaces are super extensible… for run time extensions there is a myriad of sandbox runtimes. github.com/mackross/agent… is my extensible and durable Golang agent runtime.

Pi wouldn’t make any sense in rust or go. Extensibility is key to it. That leaves ruby, python, js, php for the most part unless you want to ship an interpreter. None of those languages have any benefit over node.

@yacineMTB awkward af

@mackross Just use Google TTS like a normal person

gpt 5.5 has changed my life. my kid has been sick the past couple of days and ive been hanging out with him, but set up a tmux fork with TTS and automatic sshing to all my boxes. and man. im getting more work done than ever

@yacineMTB I really want to give the new realtime-2 as a main agent orchestrator a spin for hands-free use. Termius on my phone is straining my eyes with the teeny-tiny font I'm codexing and reviewing code on.

i'm not memeing. genuinely my life has changed. as a dad of a young family this has improved my life in a manner that is hard to describe. i'm never going to get this time back. my work getting automated is the best thing that could have happened to me

@joemasilotti @Schappi blog.cloudflare.com/code-mode/ I’ve been working on a version for local tools (in preview)

@mackross @Schappi Codemode?

MCP or CLI?