The Agent Stack
2026-02-27
by Uri Walevski
Most agent frameworks focus on the brain. The reasoning loop, the planning, the chain of thought. That part is getting commoditized fast. What's actually hard is everything around it. The hands and feet. How does your agent talk to people? How does it browse the web? How does it sign up for services? How do you watch what it's doing?
I've been building tools for this layer for a while now. They started as separate projects but ended up forming a stack. Here's what it looks like.
Supergreen -- WhatsApp and Telegram Without the Bureaucracy
If you want your agent on WhatsApp, you're supposed to go through Meta's Business API. That means applying, waiting, getting approved, dealing with message templates, rate limits, and a whole approval process that assumes you're a brand sending marketing blasts.
Supergreen skips all that. You connect a phone number, get a webhook, and start receiving and sending messages. Each account runs in its own isolated container with its own proxy. It works for Telegram too.
It's the plumbing. You don't think about it, it just works. Your agent gets a real WhatsApp presence without the enterprise sales process.
AgentMail -- Real Email for Agents
Agents need email addresses. Not throwaway ones that land in spam, real ones that can sign up for GitHub, receive verification codes, and carry on conversations with humans.
AgentMail gives each agent its own @theagentmail.net address. There's a karma system that keeps the shared domain clean. Sending costs karma. Getting replies from real people earns it back. Agents that have genuine conversations sustain themselves. Spammers run out of karma and get blocked.
You create an account, generate a scoped token for your agent, point it at the docs page, and it figures out the API on its own.
View-Chat -- See What Your Agents Are Saying
Once your agent is talking to people, you need to know what it's saying. View-chat is a conversation logger. Drop in a few API calls and every message your bot sends or receives shows up in a clean interface.
It sounds simple because it is. But without it you're flying blind. You find out your agent said something weird when a user complains, not when it happens.
Alice and Bot -- Chat Infrastructure You Actually Own
WhatsApp and Telegram are someone else's platform. You play by their rules. They can change the API, throttle you, or ban you without explanation.
Alice and Bot is a chat protocol built for bots from the ground up. End-to-end encrypted, no phone number required, embeddable as a widget on any page with one script tag. Identities are key pairs you own, not accounts on someone else's server.
It's what you use when you want a chat experience on your own terms. Embed it on your site, connect a webhook, and your agent is live.
prompt2bot -- The Whole Thing Wired Together
This is where it all converges. Prompt2bot lets you describe a bot in plain language and deploy it across all these channels: WhatsApp (via Supergreen), Telegram, email, web (via Alice and Bot), Facebook Messenger.
But the interesting parts are what the bot can actually do once it's running.
Skills. There's a community skill system where anyone can publish a skill as an npm package or a GitHub repo with a SKILL.md file. Your bot installs skills and learns new capabilities at runtime. Search the web, manage a Shopify store, schedule meetings, post to social media. If someone's built the skill, your bot can use it.
VMs with secure secrets. Your bot can spin up actual virtual machines, run code on them, and manage long-running tasks. Secrets (API keys, tokens) are handled through an encryption proxy. The VM never sees the real secret values. It gets HMAC placeholders that the proxy swaps for real credentials only on outbound requests to the right hosts. So your bot can use your GitHub token without ever having access to the actual token.
Browsers. Bots can read any URL, and when JavaScript rendering is needed, they get a real headless browser. PDF reading, YouTube transcript extraction, domain crawling for knowledge bases. The web becomes the bot's reference library.
The bot also gets hallucination detection (a fast model checks every response and auto-corrects), persistent memory across conversations, proactive messaging on schedules, and a code execution sandbox for running scripts.
Why a Stack
Each of these tools works standalone. You can use AgentMail without prompt2bot, or Supergreen without anything else. But together they cover the full surface area of what an agent needs to operate in the real world.
The brain is the easy part now. The hard part is giving it a body.
← All posts