No-Code Agentist

it's been a year since our last computer using agents piece and a lot has changed! @fabrisera2000 @zephratic and I went through and shared some data around what they can and can't do today https://t.co/mbWkO2MQSe

Even with me overseas for vacation the SMF Works AI team has been hard at work https://t.co/GiT4Vo3Uy7

There is an asymmetry in most agentic workflows that does not get talked about much: humans have many ways to talk to agents, and almost no standardized way for agents to talk back to humans. You can write a prompt. You can define a task. You can interrupt the run. But if an agent is running in the background for four hours and hits a decision point, it either blocks and waits, fails silently, or guesses and continues. None of those outcomes are what you want. Remoko is an MCP tool that gives agents an iOS push notification channel back to you. The use cases Yohei shipped with it are specific: questions, approvals, check-ins, feedback, update/cancel requests, and execution reports. These are not notifications that say "your task completed." They are structured interrupts: the agent needs something from you mid-run. This matters more as agents run longer. A Codex task that runs for 20 minutes can probably wait for you to check a browser tab. A Claude Code workflow running overnight cannot. At that length, the question of how an agent reaches a human for a non-blocking check-in becomes a real infrastructure problem. The MCP standard is what makes Remoko generalized. Any agent framework that exposes MCP tools can use it: Codex, Claude Code, or any custom agent that speaks the protocol. The notification types map to a real taxonomy of what a long-horizon agent actually needs from a human: not just "done" or "failed" but "I am at a branch point, here are my options, which do you want?" BabyAGI was one of the first public demos of multi-step autonomous agents in 2023. Three years later, the problem Remoko solves is exactly what was always going to be the bottleneck: not whether agents can run autonomously, but how humans stay in the supervisory loop when they do. https://t.co/3Ys2cWGQnM

you stopped a workflow. three agents out of four had finished one comes back Anthropic documented how it works: on resume it doesn't look at who finished, it looks at who started when agents A, B, C, D. you hit stop while B is running only A comes back from cache B, C and D all rerun, even though C and D were done what to do about it: 1 → many small agents instead of a few long ones. when it breaks, you lose less 2 → keep the first agent short. it's the point you get rolled back to 3 → don't put a human inside the run, there's nowhere to pause. need a sign-off? split it into two workflows 4 → set the ceiling up front: small is under 5 agents, medium under 15, large under 50 and the big one: resume only lives inside the session. close Claude Code and it all starts from zero, no matter how much finished

🐾 BAICLAW: AI AGENTS MADE SIMPLE, POWERFUL, AND PRODUCTION-READY No complex setup. No coding barriers. No fragmented workflows. Just install, connect, and run AI Agents through a clean visual interface. 📋 Core Positioning 🚫 No complex setup 🚫 No coding barriers 🚫 No fragmented workflows 🆓 Free, local-first, multi-platform 🧠 Breaking Down Whether "Simple" and "Production-Ready" Actually Coexist These two claims sit in real tension more often than marketing copy tends to acknowledge. Simplicity usually comes from limiting configuration and edge-case handling, while production-readiness usually requires exactly the kind of robust edge-case handling that simplicity tends to strip away. 💭 My Take A lot of no-code tools historically succeeded at being genuinely simple while quietly failing at being production-ready the moment a real-world use case exceeded the narrow set of scenarios the tool was originally tested around. The tension isn't automatically fatal to this claim, but it is the actual thing worth testing directly rather than assuming both properties coexist by default. "Local-first" is the architectural detail I'd weight most heavily as a genuine differentiator, a meaningfully different design choice than most AI Agent platforms, which tend to default to cloud-first. Local-first typically means less dependency on a specific vendor's servers staying available, and often better handling of sensitive data since it doesn't need to leave the user's own machine. Free, local-first, and multi-platform together suggest this is positioned as broadly accessible infrastructure rather than a premium tool gated behind pricing tiers, prioritizing adoption over immediate revenue. 🎯 What I'd Actually Check Whether "production-ready" genuinely holds up for complex, multi-step agent workflows, or is accurate mainly for simpler use cases and starts showing real cracks once someone pushes past basic automation. 🔗 Try it now: https://t.co/U3pq0uQ9I2 @justinsuntron @BAI_AGI #TRONEcoStar