Manzar Anjum.

Work/Production platform

The agent platform I build and support client systems with

Up to 14 agents work at once on this platform. I build and support client systems with it, and a person signs off on anything that reaches a customer or a bank account.

Role
Design, build and operations
In use
Every working day
Stack
Python, PostgreSQL and pgvector, Docker, Linux, Playwright, Claude and OpenAI with tool calling and MCP
RequestTask, written downwhat / how / proofAgent works the taskone owner, start to finishCode checksread every commandPersonapprovesbefore it runsmessages, payments,deletes, client serversSecond agent verifiescold start, live systemclaims doneVerdict + recordpass tied to the filesthe record isthe next task's memory
Fig. 01 How a task moves: written down, checked by code, approved by a person, verified by a second agentDiagram

The problem

An agent that is right most of the time is still a liability.

The one time it is wrong, it e-mails the wrong customer, or reports a job finished that never ran. A small office has nobody spare to check every step by hand, so the checking has to be part of the system.

What I built

  • Every task starts written down.

    A request becomes a task that states what is wanted, how it will be built and how it will be proven. One agent owns it from start to finish.

  • Rules the agent cannot talk its way past.

    Before any command runs, code reads it: the rule layer is 185 small programs. Sending a client message, moving money, deleting a record or changing a client's server stops there until a person approves. These are code checks, so a persuasive prompt changes nothing.

  • An independent second opinion.

    When the agent says it is done, a second agent starts cold, with no memory of the work. It tests each claim on the live system and writes a verdict. Done means passed, and the pass is tied to the exact files it checked. Change one later and the task shows as unproven again.

  • Memory that cites its sources.

    Agents remember nothing between tasks. What they know comes from the written record, searched with the RAG system I built for it, and every answer names where it came from.

How I know it works

  • A person approves every client message and every payment before it goes out.
  • A separate agent verifies each finished task on the live system: 3,285 verdicts between 12 October and 15 November 2026, and 934 of them failed the work.
  • Each morning, automated checks run through two client offices' daily workflows, 17 to 38 steps each, before staff log in.
  • When a check itself fails, it alerts a person. Nothing quietly falls back to a second path.

In numbers

  • 185small programs in the rule layer
  • 14agents in parallel at peak
  • 3,285independent checks in five weeks. 934 failed the work

My part

I designed the system and built it with AI coding agents. I deployed it and operate it day to day. No change counts as done until a separate, fresh check passes it on the running system.

  • Python
  • PostgreSQL
  • pgvector
  • Docker
  • Claude
  • OpenAI
  • MCP
  • Playwright