- unwind ai
- Posts
- Build AI Coding Agents that Run ∞ in the Cloud
Build AI Coding Agents that Run ∞ in the Cloud
+ Search, Fetch, Browser, and Web agent in one CLI + Skill
Shubham Saboo & Gargi Gupta
April 15, 2026
Today’s top AI Highlights:
& so much more!
Read time: 3 mins
AI Tutorial
I’ve been running a bunch of agents every day for months.
The problem is that I was the one constantly tweaking and learning while they just held onto context.
I tested it out by putting the same Monica agent on Hermes at the same time.
She started creating her own playbook from my edits and keeps getting better without me.
In this blog, I detailed how I did it. You’ll also understand the difference between an agent you manage and one that actually grows with you.
We share hands-on tutorials like this every week, designed to help you stay ahead in the world of AI. If you're serious about leveling up your AI skills and staying ahead of the curve, subscribe now and be the first to access our latest tutorials.
Latest Developments
One npm install and a Skill markdown file - that's all it takes to give your coding agent full access to the live web.
TinyFish just released their CLI + Agent Skill that lets Claude Code, Codex, Cursor, OpenClaw, and other agents autonomously call four web primitives directly from the terminal: Search, Fetch, Browser, and Agent.
Everything runs on their infrastructure, so you can automate workflows at scale.
You install the skill file once, and your agent figures out the rest: which endpoint to hit, how to structure the call, where to write the output. No integration code needed.
All four primitives are built in-house on TinyFish's custom Chromium engine with 28 anti-bot mechanisms at the C++ level. The platform already ranks #1 on Mind2Web (89.9% accuracy) with 40M+ operations running for clients like Google, DoorDash, and Volkswagen.
Key Highlights:
Skills over MCP for execution: The skill approach uses 87% fewer tokens per operation, writes output to the filesystem instead of the context window, and delivers 2x higher task completion on complex tasks.
End-to-end signal loop: Because TinyFish owns search, fetch, browser, and agent, every run teaches the entire stack to improve - something you can't build by assembling third-party APIs.
One platform, one session: Same IP, same fingerprint, same cookies across your whole workflow. Sites see one client, not five disconnected tools making separate requests.
Free tier available: 500 steps, no credit card required, with an open-source cookbook and skills on GitHub.
Mistral Vibe gives you a configurable agent runtime: composable skills, custom subagents, unified modes, MCP integrations, and fine-grained control over what the agent is allowed to do.
For developers who want their tools to work the way they work, this is worth a close look.
Key highlights:
Unified agent modes: Configure custom modes that combine tools, permissions, and behaviors. Switch contexts without switching tools - from a read-only exploration mode to a fully autonomous execution mode, all managed through a single config.
Custom subagents on demand: Define specialized agents for recurring tasks: deploy scripts, PR reviews, test generation. Invoke them when needed; they run independently and return results asynchronously.
Skills with .agents/skills/ standard support: Skills extend agents with new tools, slash commands, and specialized behaviors. Vibe now supports the .agents/skills/ standard, improving interoperability across agent tooling.
Web search, notifications, and session resume: The latest updates add web search mid-session, ping notifications when the agent needs attention, and /resume to continue any previous session with full context.
MCP in two steps: Add your preferred tool to config.toml and it's connected. No wrappers, no boilerplate beyond the config tweak and an AGENTS.md file in the workspace.
Vercel looked at how Stripe, Spotify, and Block build internal coding agents, and decided to open source the entire playbook.
Open Agents is a ready-to-fork platform for spinning up cloud coding agents that go from a chat prompt to committed code changes, no laptop required.
The platform separates the agent brain from its execution sandbox, which turns out to be the key architectural decision. It means your model choices, sandbox lifecycle, and agent logic can all move independently. It ships with a chat UI, durable workflow runtime, isolated VMs with git baked in, and multi-model support through AI Gateway.
Everything deploys on Vercel infrastructure and the repo is explicitly built to be customized, not consumed.
Key Highlights:
Three-Layer Design: Web app handles auth and UI, the agent runs as a durable workflow outside the sandbox, and the VM provides filesystem, shell, and git access. Clean separation means each layer scales on its own.
Workflows That Don't Die: Agent runs survive restarts, retry failures, and auto-checkpoint. You can disconnect and reconnect to a running session from anywhere.
Sandbox-Per-Session: Every coding session gets an isolated environment with repo cloning, branch management, auto-commit, and optional PR creation. Sandboxes hibernate on inactivity and resume from snapshots.
Built to Fork: Not a SaaS product, it's a template. Deploy your own instance, wire up your GitHub App, swap in your preferred models, and build your team's custom coding agent platform.
Quick Bites
Run your Claude Code Routines on autopilot
Anthropic just shipped "Routines" for Claude Code. Configure a prompt, point it at a repo and your connectors, then let it run on a schedule, via API webhook, or in response to GitHub events, all on Anthropic's cloud so your laptop can stay shut. Think nightly issue triage, auto-porting PRs across SDKs, or Datadog alerts that arrive pre-triaged with a draft fix. Available today across all paid plans, with Claude Code on the web enabled.
What really is an Agent Harness?
Theo's latest video is the most fun explanation of AI coding harnesses you'll find — he builds one in ~60 lines of Python, then proceeds to lie to the model about what its tools do (it just believes you), swap the entire toolset for a single bash command (the model figures out file I/O on its own), and show why the same Opus model jumps from 77% to 93% just by moving it into Cursor. His core argument: the harness is the product, and whoever obsessively tweaks their system prompts and tool descriptions per model wins. Do watch this.
Turn your AI prompts in Chrome into reusable Skills
Chrome now lets you turn your best Gemini prompts into saved Skills - reusable, one-click AI workflows you can fire on any page or across multiple tabs. Great abstraction: instead of re-typing the same prompt every time you hit a recipe or a product page, you save it once, customize it, and trigger it with / going forward. Google's also shipping a curated Skills library with ready-made options, and your saved Skills sync across all signed-in Chrome desktops.
Claude Code's new desktop app is built for parallel agents
Anthropic just dropped a major redesign of the Claude Code desktop app, and it's clearly built for the multi-agent workflow most of us have been using - spinning up parallel sessions across repos, each doing its own thing while you play air traffic controller. The new sidebar tracks all active sessions, there's a drag-and-drop layout with an integrated terminal and file editor, and side chats let you branch off quick questions without polluting the main task context. Available now on Pro, Max, Team, and Enterprise.
Tools of the Trade
Self-Improving Agent Skills: Automatically optimize your agent skills using a multi-agent system built with Google ADK (Agent Development Kit) and Gemini. Upload a skill, let the agents generate test scenarios and evaluation criteria, then watch as three specialized ADK agents collaborate to improve your skill through iterative optimization.
Kontext CLI: Open-source tool that wraps AI coding agents with enterprise-grade identity, credential management, and governance. Instead of copy-pasting long-lived API keys, you declare dependencies, and the CLI handles OIDC auth, short-lived token exchange, and automatic expiry.
Claude Reflect: A Claude Code plugin that builds persistent memory by automatically capturing your corrections and preferences during coding sessions, then syncing them to CLAUDE.md so Claude doesn't repeat the same mistakes. It also analyzes your session history to spot repeating workflows and generates reusable slash commands from them.
Awesome LLM Apps - A curated collection of LLM apps with RAG, AI Agents, multi-agent teams, MCP, voice agents, and more. The apps use models from OpenAI, Anthropic, Google, and open-source models like DeepSeek, Qwen, and Llama that you can run locally on your computer.
(Now accepting GitHub sponsorships)
Hot Takes
Markdown is the new Python
the ai labs, in competing with each other, are burning huge amounts of the commons on public trust in ai to win minor points against the others. their lobbyists, pr machines, lawsuits. it’s the very opposite of what marxist class struggle analysis would tell you
~ roon
That’s all for today! See you tomorrow with more such AI-filled content.
Don’t forget to share this newsletter on your social channels and tag Unwind AI to support us!
PS: We curate this AI newsletter every day for FREE, your support is what keeps us going. If you find value in what you read, share it with at least one, two (or 20) of your friends 😉
Reply