The AI Spotlight Follow
One Prompt, Three Hours: A Builder Fused the Best Parts of Six AI Agent Apps Into One Platform He Owns
AI Agents

One Prompt, Three Hours: A Builder Fused the Best Parts of Six AI Agent Apps Into One Platform He Owns

Grok Bot's interface, Dots' computer, Muse's social reach, iMessage from Instinct and the open-source Hermes harness underneath. One builder listed what he loved about every agent platform, wrote one long prompt for Claude Fable 5.2, and came back three hours later to a working app. Here's the teardown, the exact recipe and the caveats.

Advertisement

Every AI company now sells an "agent platform." Each one does something brilliantly and something else badly.

So one builder tried an experiment. He sat down with six of them, wrote down exactly what he loved about each, and handed the whole wish list to an AI coding model in one prompt.

Three hours later, he had a working app: Grok Bot's clean interface, a computer that looks like OpenAI's, messages over iMessage, a real-time voice that can drive a browser, and an open-source engine underneath that runs on any model he chooses.

He didn't touch the code. He just sent the prompt and came back.

Advertisement

Here's how it worked, step by step, and what it tells you about building with agents right now.

Step one: tear down the competition

The first half of the project wasn't coding. It was product research. He walked through each platform and noted one or two things worth stealing.

Grok Bot: the interface and the voice

Grok Bot's strength is simplicity. Spinning up a bot is easy, bots can talk to each other, and, crucially, each session is a persistent bot identity dedicated to one job, not just another chat window.

Connecting apps is simple: add to Slack, create a routine, see the bot's computer. There's voice chat with selectable voices and speeds, and a library of every asset a bot has produced.

His gripes: the bot's computer can be slow to load. His praise: messages feel like texts. Bots rarely send walls of text, so replies are quick to digest, which matters if you have ADHD and want to scan and move on.

He also rated xAI's voice models highly, especially for tool use in real-time voice, and singled out Grok Bot's native X (Twitter) search for research on what's going viral.

Stolen: the UI/UX, the voice, and X research.

Muse: the social reach

Muse is Meta's agent. It uses the same iMessage-style texting format, but it's one bot, not a team. If you want to multitask, you open more chats.

What Muse does uniquely well is social: native connections to Instagram, Facebook and Threads, so it can help reply to DMs or find potential clients engaging with your content. It exposes editable "soul" and memory files, an idea its makers have said was inspired by the open-source OpenClaw project. It also has a personalized digest feed and nudges you with use-case ideas, a sign it's aimed at consumers.

His summary: Muse is a personal agent. Grok Bot is a work agent.

Stolen: social capabilities, and the idea of editable soul and memory files.

Dots: the computer

Dots is OpenAI's agent platform, launched in late September. Again, a single agent. It can connect to Slack and your devices and manage multiple Codex coding sessions behind the scenes.

His favorite feature: the agent's computer. It feels like the agent has its own home, with apps installed and a smooth hand-off when you take control. He also liked a subtle "thinking" indicator in the interface.

Stolen: the computer's look and feel, the thinking UI, and quiet multi-session management.

Hermes: the engine

Hermes Agent, from Nous Research, is an open-source agent harness. It followed OpenClaw and, in his view, overtook it by being more reliable: it doesn't break on its own.

Hermes is MIT-licensed, model-agnostic, and supports persistent memory, reusable skills, tools, MCP and subagents. You can plug in your existing model subscriptions or API keys.

His complaint: the desktop app's design. Busy sidebar, an attempt at Grok Bot-style bots that isn't quite right, a screen you have to start manually.

His verdict: the best agent harness available, wrapped in the least polished interface.

Stolen: the entire engine.

Instinct and Duet

Instinct is an iMessage-based personal agent. The one thing he wanted from it: the iMessage bubble.

Claude Duet is Anthropic's answer to Muse and Dots, announced in October as a waitlisted research preview for paid US subscribers. He hadn't gotten access yet, but he wanted Anthropic's models, Claude Opus 5.5 and the newly released Claude Fable 5.2, inside his build.

The design choice hiding underneath: one bot or many?

The teardown surfaced a real split in how companies think about agents.

Single-agent platforms (Muse, Dots, Duet, Instinct) give you one assistant that knows everything about you. The upside: one memory, one relationship, nothing to configure. The downside: everything piles into one identity. Multitasking means juggling chats, and a personal assistant that also runs your sales pipeline can get confused about context.

Multi-agent platforms (Grok Bot, and Hermes' attempt at it) give each job its own persistent bot. The upside: clean separation, a dedicated bot for research, another for support, each with its own instructions and history, and they can message each other. The downside: more to manage, and the temptation to create dozens of bots you never maintain.

His build split the difference. Up to four or five named bots for distinct roles, with everything else handled as threads and sessions inside them. Plus one shared computer with a screen per bot, which keeps costs down compared with a separate machine for every agent.

That's a sensible default for most small teams: a handful of specialists, not an army.

Step two: source every piece before prompting

This is the part most people skip, and it's why the build worked.

Before writing the prompt, he mapped each wanted feature to an existing component, and color-coded what he knew he could get (green) versus what was uncertain (yellow).

Feature Where it came from
Grok Bot-style interface His own earlier open-source clone of the Grok Bot UI, built by having an AI click through the app and redesign it in a design tool via MCP
Voice xAI's voice models
X research X's API and MCP server
Instagram social features Uncertain: no clean public DM API, so flagged yellow, with a possible workaround via an integration platform
Dots-style computer A cloud computer, styled using his own earlier open-source project that mimicked the Dots look
Agent engine A fork of the Hermes Agent repo
iMessage A third-party iMessage provider; his Hermes agent already had a phone number
App connections and logins Composio for authentication and connectors
Email inbox for the agent AgentMail
Long-term memory Honcho, which he found cheapest and most durable after trying alternatives

His rule: find something on GitHub that's already 90% there, give the repo to your coding agent, and let it rip.

Step three: the one prompt

He worked inside T3 Code, a free, open-source agent development environment that manages coding agents across multiple machines, with your own subscriptions. He connected cloud machines for Linux, Mac and Windows so agents could test on every OS, and chose a Linux box for this build.

Model: Claude Fable 5.2, set first to medium effort, then bumped to high "just to be safe."

Then he dictated the prompt by voice. Its key instructions, in order:

  1. Build my own agent bot platform with Grok Bot's design and UI, reusing the clone I already built.
  2. Keep the xAI voice modes we already implemented.
  3. Give agents X research capabilities, using X's API and MCP.
  4. Make the computer look like Dots', using my earlier repo as the reference (attached).
  5. One screen per bot, not five. Use a single cloud computer for all bots, with a separate screen for each, like Grok Bot does. Support up to four or five bots; everything else becomes threads and sessions.
  6. Copy the Dots thinking UI (screenshot attached).
  7. Use Hermes as the engine and harness for every bot (repo linked). Either wrap Hermes' Electron desktop app or rebuild from the ground up, "whatever you think is best."
  8. Preserve model switching, so existing Claude, OpenAI and Grok subscriptions work.
  9. Use my existing Hermes agent (already on iMessage) as the underlying agent, and restyle its computer to match.
  10. Wire in Composio for auth and connectors, alongside Honcho memory and AgentMail.
  11. Think about all the moving parts and blend them into a platform "that I own."

Then he hit send and walked away.

Step four: three hours later

When he came back, that single message was still the only one he had sent. Fable 5.2 had worked for three hours on high effort.

It had done something he didn't ask for: it hosted the app on the cloud computer, exposed a public URL, sent back screenshots, and started testing the app itself.

What worked

  • Login through Composio.
  • The interface: a faithful Grok Bot-style layout, with messages already in the thread.
  • X search was implemented and working.
  • The computer panel, styled like Dots, with take-control and hand-back.
  • Add to Slack via Composio.
  • The engine: asked what harness it was running, the bot answered Hermes Agent harness, running Claude Opus 5.5 via the API.

Bots talking to bots

He created a second bot and asked it to ask the first bot what it knew about him. It messaged the first bot, got a reply and relayed it, complete with personal details he had to blur. Each bot had its own screen on the shared computer, as requested. He found the computer noticeably smoother and faster than the one in the app he'd copied.

iMessage and voice

He texted the agent's number from his phone, then asked the agent in the web app what he'd just texted. It answered correctly. Same agent, same memory, two channels.

Then voice. He asked it to open a chess site and his YouTube channel. It opened both in separate tabs, switched when asked, and sent a message to the other bot on request, all by voice, quickly.

His reaction, caught in the moment, was that it felt like AGI. That's enthusiasm, not a technical claim. But the result was undeniably a working product built from one prompt.

He said he planned to open-source the result.

A starter template you can adapt

You don't need his exact stack to copy the method. Here's a skeleton prompt structure based on what worked, to fill in with your own components:

  1. Goal: "Build an agent platform I own that combines X, Y and Z."
  2. Interface reference: "Match the design of [repo or screenshots attached]."
  3. Engine: "Use [open-source harness, repo linked] for every bot. Preserve the ability to switch between my existing model subscriptions."
  4. Computer: "Give bots a computer styled like [reference]. One shared machine, one screen per bot."
  5. Channels: "Bots must be reachable via [Slack / email / iMessage provider]."
  6. Services: "Use [auth/connector service] for logins, [memory service] for long-term memory, [email service] for inboxes."
  7. Limits: "Support up to N bots. Everything else is threads."
  8. Freedom: "Choose whether to wrap the existing app or rebuild. Explain your choice."
  9. Verification: "Host it, test each feature end to end, and send me screenshots and a list of anything that failed."

The last line matters most. Asking the model to test and report its own failures turns a long unattended run into something you can actually review.

What it costs to own

Owning your platform means owning its upkeep. Expect to update the harness as it ships changes (Hermes' maintainers commit constantly), rotate API keys, watch the bills for each service, and patch security issues. For a solo builder, that's a few hours a month. For a company, it needs an owner.

Why this worked when most one-shot builds fail

Strip away the excitement and the recipe is replicable:

  1. Research before you prompt. He knew exactly what he wanted and why.
  2. Reuse, don't reinvent. Nearly every component already existed as a repo, API or service. The model's job was integration.
  3. Attach references. Repos and screenshots gave the model concrete targets rather than vague descriptions.
  4. State constraints. "One screen per bot." "Four or five bots max." Constraints prevent sprawl.
  5. Use a strong model at high effort, and give it time. Long-horizon work needs long runs.
  6. Let it test itself. Hosting and self-testing caught issues before he looked.

Why this matters

For two years, the question was which company would own the agent platform. This build hints at another answer: maybe nobody has to.

When harnesses like Hermes are open source, connectors and memory are services, and frontier models can integrate them in an afternoon, the "platform" becomes something you assemble. Your bots, your models, your data, your interface.

That's good for users who hate lock-in and usage caps. It's a challenge for companies whose main moat is a nice interface.

Practical takeaways

  • Write your feature wish list from real products, one or two items each, before touching code.
  • Map every feature to an existing component and flag the uncertain ones honestly.
  • Pick a model-agnostic harness so you can switch providers as prices and quality change.
  • Use managed services for the hard plumbing: auth, email, memory.
  • Constrain scope in the prompt to keep the build manageable.
  • Budget for usage. A three-hour high-effort run on a frontier model isn't free.

Counterpoints and risks

  • Demo versus production. A working prototype isn't a hardened product. Security review, error handling and maintenance still need humans.
  • Security surface. Agents with iMessage, email, Slack, a computer and app logins hold the keys to your digital life. Isolate them, scope permissions and log actions.
  • Platform terms. Wrapping or cloning commercial apps' designs, or automating social platforms without official APIs, can breach terms of service.
  • Hidden dependencies. The build leaned on pre-existing repos and an already-configured agent. Starting from zero takes longer.
  • Cost. Long frontier-model runs and multiple paid services add up.

FAQ

Can you really build an agent platform with one prompt? In this case, yes, but only because nearly every component already existed and was specified clearly. The prompt was long and detailed, and the model ran for three hours.

What is Hermes Agent? An MIT-licensed, open-source agent harness from Nous Research with memory, skills, tools, MCP support and subagents. It works with many model providers.

What model built it? Claude Fable 5.2 on high effort, inside the T3 Code environment. The finished app's bots ran on Claude Opus 5.5 through Hermes.

Do I need to know how to code? Less than before, but you need to know what you want, find the right components and review the result for security.

The bottom line

The best agent platform might not be one you buy. It might be the one you assemble from the best parts of everyone else's.

One builder proved the point with a single prompt: research first, reuse everything, constrain the scope, pick a strong model and give it time.

Three hours. One message. A platform he owns.

Sources

Someone Built a Free, Open-Source Grok Bot. It's One of the GitHub Repos Agent Builders Are Grabbing Right NowUp next · 11 min readSomeone Built a Free, Open-Source Grok Bot. It's One of the GitHub Repos Agent Builders Are Grabbing Right NowOpenMausBot gives you Grok Bot-style teams of AI bots that run on your own subscriptions. It's one of a wave of trending repos, from Anthropic's open skills and a username finder that checks 400 sites to Nvidia's data ontology tool and DeepMind's MuJoCo 4.0. Here's what each does, who it's for, and the catches.Keep reading →