Claude Code, MCP, and building AI agents with human-in-the-loop workflows.
A new per-workspace setting holds outgoing chat messages for a few seconds before delivery, with Send Now and Cancel controls — a brief buffer to catch a typo or reconsider before an agent acts on it.
Read PostTasks created in the wrong workspace used to mean starting over. Right-click a task, pick a destination, and it moves with its full chat history, tool calls, and attachments intact.
Read PostWe ran an independent MCP conformance suite against our Core and Workspace MCP servers, fixed the real bugs it found, and wrote tests pinning down the cases where it was simply wrong. Here's what that looked like in practice.
Read PostAgentRQ's Slack integration turns a workspace into a shared room: one private channel, every teammate's own agent, and every approve/deny attributed to the person who clicked it. Here's how team collaboration works without anyone touching a shared login.
Read PostAgentRQ workflows have no if/else, no conditions, no decision nodes. Routing is a lookup on (workflow, event), branching is a choice of event name, and loops are rejected at save time. Here's the mechanism, tested against a live v0.4.1 build.
Read PostAgentRQ v0.4.0 adds Workflows — chain events and workspaces into one named pipeline you can draw or type, review before it runs, and execute in isolation. Agents coordinate by publishing named events: no conditionals, no back-and-forth.
Read PostA new AgentRQ plugin for DeepSeek Harness lets you assign tasks from the AgentRQ web app or your phone and have them pushed live into a running DeepSeek Harness agent — no polling, no leaving the harness.
Read PostEvery tool call and message on a task now lives in one searchable trajectory view — a 3-lane minimap plus a master-detail panel with Summary, Payload, and Content tabs.
Read PostAgentRQ's workspace analytics page now includes GitHub-style contribution heatmaps for task completions and messages, adapting from hour-by-hour to day-by-day as you zoom out.
Read PostSpeak your task instructions instead of typing them — AgentRQ transcribes speech to text with a local Whisper model running in your browser, no audio ever leaves your device.
Read PostAgentRQ can now write your task titles for you — using a small language model that runs entirely inside your browser, so the task description never leaves your machine.
Read PostStep-by-step guide on running Gemma 4 31B with Multi-Token Prediction (MTP) on Nvidia NGX Spark using Docker and Claude Code.
Read PostAnthropic has unlocked Claude Channels for all Claude Code users, removing the Claude Pro/Max subscription requirement. Here is what this means for API-based developers and the open-source community.
Read PostAnnouncing the AgentRQ Codex Gateway. Connect OpenAI Codex and other ACP-compatible agents to your AgentRQ workspaces for remote human-in-the-loop task management.
Read PostDiscover how Agentic Task Management bridges the gap between autonomous AI agents and human oversight using AgentRQ.
Read PostAgentRQ's two-tier MCP architecture — a supervisor with global visibility and workers scoped to individual workspaces — is what makes closed-loop, self-growing multi-agent systems actually work in production. Here's how and why.
Read PostClosed-loop agentic systems continuously improve through feedback cycles. Open-loop systems execute and forget. Here's why the feedback loop is the single biggest determinant of which AI platforms win — and how AgentRQ is built around it.
Read PostDeepSeek-V4-Pro is a 1.6 trillion parameter open-source MoE model with MIT license, 1M token context, and benchmark results that challenge the best closed-source models. Here's what it actually is, how it works, and what it can do.
Read PostLearn how to connect Gemini CLI and other ACP-compatible agents to AgentRQ using the new open-source ACP Gateway. Remote task management for your local AI agents.
Read PostMCP Gateways solve the hardest problems in connecting AI agents to real APIs: authentication, security, context overhead, and observability. Here's what they do and how HasMCP implements them.
Read PostAgentic workflows and autonomous workflows aren't the same thing. Here's how they differ, when each applies, and why the distinction matters for teams building with AI agents.
Read PostAnthropic's unreleased Claude Mythos model has discovered thousands of zero-day vulnerabilities autonomously. Here is what we know about Mythos, its defensive role in Project Glasswing, and the future of frontier AI cybersecurity.
Read PostGoogle's Gemma 4 now runs entirely on iPhone via the AI Edge Gallery app — with Thinking Mode and Agent Skills, fully offline. Here's why it matters for mobile AI, device control, and private intelligence.
Read PostAn always-on agent with workspace context outperforms an on-demand zero-context agent in almost every real project scenario. Here's why — and when each approach actually makes sense.
Read PostM5 Max MacBook Pro with 128GB + Qwen3.5-35B + MolmoWeb = a self-contained AI stack that runs entirely offline. Here's what becomes possible when frontier-grade intelligence fits in a backpack.
Read PostAllen AI's MolmoWeb is a fully open 4B/8B web agent that navigates browsers using only screenshots — no DOM access, Apache 2.0 license. Here's how it works, what it actually beats, and how to run it.
Read PostGoogle's TurboQuant compresses LLM key-value caches to 3 bits with no accuracy loss — 8x faster on H100s. Here's how it works and what it means for running large models locally.
Read PostMCP Tasks is an experimental spec that turns synchronous tool calls into durable state machines — poll for status, retrieve results later, and handle mid-task human input. Here's what it means for long-running AI agents.
Read PostACP (Agent Client Protocol) is the open standard that turns any coding agent into a first-class IDE citizen. Here's what it enables today with stdio, and what becomes possible when remote transport arrives.
Read PostClaude Code notification channels flip the model from pull to push — here's what that unlocks, what the official channels cover, and how to think about what's worth wiring up.
Read PostClaude Computer Control lets AI agents interact with screens, keyboards, and mice. Here's the technical breakdown of how it works, which models support it, and why it matters for agentic workflows.
Read PostHow I use isolated AgentRQ workspaces for marketing, cold outreach, and coding — each scaling itself with human-in-the-loop workflow automations while I sleep.
Read PostMost engineers have reached Level 3. Here's the full stack of AI leverage — from chatbots to fully autonomous, human-in-the-loop agentic workflows — and what it takes to reach the top.
Read PostClaude Code Channels let you push webhooks, chat messages, and alerts into a running Claude Code session via MCP — here's how they work and how to build your own.
Read Post