A series of recent releases introduced several incremental and some more substantial updates across the AI agent ecosystem: Hermes Agent gained a Kanban-based multi-agent coordinator, Mistral Vibe added remote agents powered by Mistral Medium 3.5, and multiple projects published optimizations aimed at reducing token usage and costs. Lightweight approaches for browser control and autonomous deployments were also released.
Hermes Agent: Kanban for coordinating multiple agents
The latest Hermes Agent version ships a Kanban-style multi-agent coordination system. In this setup, individual agents have their own tools, skills, and personality profiles; they claim tasks from a board, fan out work across linked dependencies, and pass files through shared workspaces or git worktrees.
Key features:
- live dashboard for tracking progress
- per-task comment threads writable by both humans and agents
- heartbeat monitoring for job health
- SQLite-backed persistence so state survives crashes and reboots
This configuration targets organized workflows involving multiple agents where visibility and durable state management matter.
Mistral Vibe: Remote Agents and Mistral Medium 3.5
Mistral Vibe’s update introduces remote agents: asynchronous coding sessions that run in the cloud, execute in parallel, and notify you when they finish. These agents run on Mistral Medium 3.5, a 128 billion parameter (128B) dense model intended for longer-horizon, agentic tasks.
Highlights:
- remote agents can be spawned from the Vibe CLI or Le Chat UI
- local sessions can be teleported to cloud sandboxes mid-task, preserving history, state, and approvals
- Mistral Medium 3.5 weights are released under a modified MIT license and can be self-hosted; the team notes it can run on as few as four GPUs
- Le Chat includes a new "Work mode" that handles multi-step tasks (inbox triage, research synthesis, cross-tool workflows) with connectors on by default and explicit approvals for sensitive actions
Vibe’s CLI remains an open-source, terminal-native coding agent supporting slash-command skills and custom subagents.
Claude Code: fewer tokens, same technical content (Caveman skill)
A Claude Code skill and plugin called Caveman reduces response token counts by roughly 75% by stripping filler words, pleasantries, and hedging while preserving technical accuracy, code blocks, and exact error messages.
Usage:
- install: claude install-skill JuliusBrussee/caveman
- toggle on demand with /caveman or "caveman mode" and return to normal mode when needed
The intent is faster responses and lower API costs through token savings.
Browser control with under 600 lines: Browser Harness and Browser Use Desktop
Browser Harness is an open-source ~592-line project that gives an LLM a raw CDP websocket connection to Chrome, a small set of helpers, and a skill file. The agent sees how its tools work at the protocol level, writes tools when necessary, and self-corrects on failures.
The project was packaged as Browser Use Desktop, an open-source desktop app that lets Claude Code, Codex, or other coding agents control your browser without replacing Chrome. An example from the team shows an agent writing its own upload tool when it lacked one and continuing the task.
Agents can even be triggered by messaging yourself (for example, texting @BU on WhatsApp).
Autonomous account, domain purchase and deployment: Cloudflare + Stripe
Cloudflare and Stripe announced a collaboration enabling coding agents to autonomously create a Cloudflare account, buy a domain, start a paid subscription, and deploy to production without human intervention. Stripe handles identity and payment tokenization with a per-provider spending cap (noted as $100/month per provider) while Cloudflare auto-provisions accounts. To use it developers install Stripe CLI plus the Stripe Projects plugin; an open protocol is planned for release.
Cost-saving and supporting tools
- Deepclaude: a proxy that routes Claude Code’s backend to DeepSeek V4 Pro, OpenRouter, or other Anthropic-compatible APIs, preserving Claude Code’s agent loop and UX while offering roughly 17× lower cost.
- Agent-desktop: a Rust-based CLI that reads OS accessibility trees and outputs structured JSON with stable element references, enabling native app automation.
- Heard: a voice companion that reads AI coding agent terminal replies aloud via TTS while Wispr handles what you say to the agent.
- Awesome LLM Apps: a curated collection of LLM apps covering RAG, AI agents, multi-agent teams, MCP, voice agents, and models from OpenAI, Anthropic, Google, and open-source models like DeepSeek, Qwen, and Llama.
Tutorial: Anatomy of Agent SKILLS
An accompanying tutorial explains how to author Agent SKILLs effectively. The authors emphasize that the first two lines of SKILL.md are critical because they help the LLM route queries to the right skill without embeddings or retrieval layers. The tutorial breaks skills into five parts and encourages readers to implement and ship a skill for a weekly workflow.
Short summary
Recent releases focus on more efficient, cost-aware agents and simpler integrations: remote agents, token-reduction skills, compact browser control, and automated deployment flows. Hermes Agent’s Kanban feature is aimed at structured multi-agent workflows, while Mistral, Claude Code, and the various tools (Deepclaude, Agent-desktop) push toward operational simplicity and lower running costs.
Read the tutorials to try building and shipping a skill or to test any of the new tools mentioned.
