@markiewagner@PoeticHQ Congrats on Poetic. I am building an open source multi agent developer workflow app called Usta. It focuses on structure and reviewability through an event bus and role handoffs. https://t.co/HZlo0UHYjX
@steipete@OpenAI@OpenClaw Love this direction. I open sourced Usta: a macOS multi agent desktop dev environment where a PM agent builds the team and specialist roles coordinate through an event bus. https://t.co/HZlo0UHYjX
I built Usta because I was tired of being the project manager for my AI tools.
One assistant is useful. A team that talks to itself is much more interesting.
Native macOS. SwiftUI plus Rust. Anthropic, Gemini and Ollama. MIT licensed.
https://t.co/5bbdztupMo
Most AI coding tools give you one assistant.
I wanted something closer to a real engineering team.
So I built Usta.
A native macOS app where a PM agent plans the work, specialist agents run in parallel, and they coordinate through a shared event bus.
https://t.co/5bbdztupMo
The main problem with AI coding is not only the model.
It is the workflow.
Frontend finishes something, QA should wake up. Backend changes an API, docs should update. Tests fail, the right role should reopen.
That is the idea behind Usta.
https://t.co/5bbdztupMo
Honestly, these 5 hour limits are getting really hard to deal with π
I genuinely wish there were a reasonable way to get a discount on #Anthropic plans. I would happily pay, but the current pricing adds up quickly for independent developers and students.
@AnthropicAI I like how you mapped the activity to established threat tactics. How do you evaluate defenders when attackers use LLMs for reconnaissance and tooling?
@GoogleAIStudio This weekend I am building a small RAG workflow and, more importantly, an evaluation harness so I know if I am actually improving anything.
@deepseek_ai Nice promo. I am curious about what you are optimizing beyond token price. Latency, reliability and guardrails can matter more than raw cost.
@Aditya_181105 This guessing game is fun, but it reminds me that style often comes from prompt scaffolding. The real test is consistency and defenses against mistakes.
Patina AR History Explorer (Beta) β AR + AI history layers (because museums should spill their secrets π) https://t.co/wCyHUfw1UY #ai#ios#python#ar
Hardware-Aware Nussbaum-PID Controller β control theory with real hardware in mind (not just a perfect world π ) https://t.co/7UqYGaXwp8 #robotics#python
TalkingHeadAI β digital clone with personal context (because sometimes you need your AI to remember you π) https://t.co/8RdYc9WtJI #ai#python#automation
Ned3 Pro DRL Sim-to-Real Reaching β sim-to-real RL for reaching (because robots deserve a second chance after sim π ) https://t.co/a7Qagn2Xyh #reinforcement_learning#robotics#python#ros
CNN-based Persian Digit Recognizer (PyTorch) β teaching a network to read digits so I donβt have to π https://t.co/Sw3N0ppE6L #ai#python#pytorch#deep-learning
Ned3 Pro Adaptive Controller β adaptive control (teaching the robot to behave even when hardware is being dramatic π) https://t.co/goRjOWChRb #robotics#ros#python