Mirai

What Mirai does

• Drafts posts from news. A Grok bot sends stories in; Mirai reads the source and writes a ready-to-publish post.
• Asks before anything goes out. Every draft arrives as a review card. I reply to change it, or approve it and tap where it should be posted.
• Helps with the newsletter. It searches every stored story and drafts the biweekly MasterAI newsletter.
• Makes images. Headers, carousel slides and edits, in the MasterAI style, at the right size for each platform.

How I make agents

Every agent I build is four pieces: a home, a brain, hands and a door. Mirai runs on Cloudflare with Pi Durable, and these are the same four pieces in about 1,200 lines. Start with one tool and one channel, then grow it.

1. A home: one Durable Object per conversation

   A Durable Object is a tiny server with its own SQLite database. Name it after the chat (getByName(chatId)) and every conversation gets its own memory, isolated from the rest.

   JSONC

   // wrangler.jsonc
   "durable_objects": { "bindings": [{ "name": "CHAT", "class_name": "MyAgent" }] },
   "migrations": [{ "tag": "v1", "new_sqlite_classes": ["MyAgent"] }]

2. A brain: Pi Durable through PiHarness

   PiHarness runs the agent loop (model call, tool calls, next model call) and saves every step before showing it. If Cloudflare restarts the object mid-answer, the run continues where it stopped. You pick the model; I use OpenRouter.

   TS

   // bun add agents @earendil-works/pi-durable @earendil-works/pi-ai @earendil-works/chord
   export class MyAgent extends DurableObject<Env> {
     registry = createRegistry();
     harness = new PiHarness({
       harness: ({ storage, context, onWake }) => {
         this.registry.install({ name: 'mine', sections: [prompt], tools: [getNews] });
         const models = createModels();
         models.setProvider(openrouterProvider());
         return Harness.open(storage, { models, registry: this.registry, onWake }, context);
       },
       defaults: { model: { provider: 'openrouter', id: '~anthropic/claude-fable-latest' } },
     });
     lifecycle = Lifecycle.install(this).use(this.harness);
   }

3. Hands: tools

   A tool is a name, a one-line description, a typed input schema and a function. The model reads the description to decide when to call it, so write it for the model. Mark a tool replay: 'safe' only if running it twice does no harm; anything that posts or spends money stays unsafe.

   TS

   const getNews: ToolRegistration<typeof Input> = {
     name: 'get_news',
     description: 'Stored AI news stories, newest first.',
     parameters: Type.Object({ sinceDaysAgo: Type.Optional(Type.Number()) }),
     replay: 'safe',
     async execute({ sinceDaysAgo }) {
       const rows = await db.listNews({ sinceDaysAgo });
       return { output: [{ type: 'text', text: JSON.stringify(rows) }] };
     },
   };

4. A door: how messages get in and answers get out

   A Worker route receives the message (a Telegram webhook, an API call) and hands it to the right Durable Object. Use the incoming event's id as the operation id, so a retried webhook never runs twice. Answers stream back from session.events().

   TS

   // Worker: POST /telegram
   const chat = env.CHAT.getByName(String(chatId));
   await chat.receive(text, update.update_id);
   
   // Inside MyAgent
   await this.harness.submit(text, { operationId: `tg:${updateId}` });

5. Ship it

   Keys go in Doppler and reach the Worker with wrangler secret. Deploy with wrangler deploy, test with a fake event, then point the real channel at the Worker URL.

The rule I follow most: anything plain code can do stays in code. Mirai's buttons, permissions and posting never touch the model, so they're instant, free and predictable. The model gets the judgment work: reading, writing, choosing.