← All workCase study 1 of 4
AI agentOwn tool · in daily use

Career & Sales Agent

An autonomous agent that finds remote jobs and client leads, writes only from facts I've verified, and runs sales conversations on WhatsApp, Messenger and email, while every message passes a guardrail written in code and I stay one tap away.

Role
Solo: design, build, deploy, run
Status
Running 24/7 on a Linux server
Tests
281 automated
Timeline
October 2026 – now
The problem

Job hunting and client hunting are two full-time jobs.

Every day means reading hundreds of job posts and group posts, checking who can actually hire from Bangladesh, tailoring a CV, and following up with people who asked for exactly what I build. Most automation tools solve this by spamming, which burns accounts and reputation.

I wanted the opposite: an agent that does the work at volume but never sends anything I wouldn't, and never puts my accounts at risk.

What it does

Find. Write. Talk. Close.

  1. 01Find jobs and clients

    Job boards and employers' own Greenhouse, Lever and Ashby boards; Facebook groups, Google Maps and web search for people asking for software.

  2. 02Write only from verified facts

    CVs and cover letters drawn from a master CV, with a citation check that drops any line it can't trace back to a fact.

  3. 03Talk like a sales partner

    Contacts people where their post asked, qualifies need, budget and timeline in the conversation, and spots fakes.

  4. 04Close, with my price

    Proposals wait for my price on Telegram; every deal ends as won or lost with the reason, so the agent learns.

Architecture

Every action goes through one gate.

Missions15 scheduled jobs: scout, inbox, social scan, lead hunt, nurture, reports
Claude Agent SDKOpus for writing and judgement, Haiku for bulk scoring
Knowledge baseMaster CV, playbooks and lessons learned, as Markdown
Guardrail gateAutonomy level per action · daily limits · do-not-contact · blocked placeholders · link rules · 30-minute stop window · platform pause on any security check · audit log
EmailGmail API and the outreach mailbox
WhatsAppOfficial Cloud API, approved templates
Messenger · InstagramThe Page API and the agent's own browser
Job formsPlaywright fills them; the first three wait for me
ControlTelegram bot with Approve, Stop and /price · an installable dashboard on a private network
Runs onUbuntu on a home server: systemd services, a virtual screen for the browser
Missions call Claude, every outward action passes the guardrail gate, then reaches the channels. I control it from Telegram and the dashboard.
Key decisions

Where the engineering went.

  1. 01

    Guardrails live in code, not in the prompt

    A prompt can be talked around. One function decides every send: level, limit, recipient, attachments, placeholders. The model never sees a way past it.

  2. 02

    Strangers' text is data, never instructions

    Every job post, email and group post is fenced before it reaches Claude, so a hidden "ignore your rules" can't steer the agent.

  3. 03

    Never send twice

    If a message was typed but its arrival couldn't be confirmed, it counts as sent and I'm asked to check, instead of retrying into a duplicate.

  4. 04

    Moving conversations to official APIs

    Automating my personal Facebook hit encryption PINs, message requests and security checks. I moved conversations to the WhatsApp Cloud API and the Page inbox, where an agent may talk on its own.

Where it stands

Running, tested, honest about limits.

automated tests, run before every deploy
281
missions running unattended on a server
15
gate every outgoing message passes
1

Results are measured in my own pipeline, not a client's: no revenue or conversion figures are claimed.

Screenshots

Dashboard, deal board and Telegram approvals, shown with sample data on request.

Python 3.12Claude Agent SDKFastAPISQLitePlaywrightWhatsApp Cloud APITelegram Bot APIReact + Vite PWAsystemd · Tailscale