Back to home
All four

Multi-project operations · agent collaboration · permission control

Multi-project agent operating system

Keeps project state current, ranks the next actions, and sends approved work to execution agents.

Running daily

Operational review

Signal → cause → change

2 findings
  1. 13 → 0
    Signal

    The morning brief was correct thirteen times and read zero times.

    Root cause

    The system solved visibility, but required a deliberate click to deliver value.

    Change

    I replaced feature-first thinking with a zero-click delivery rule.

  2. 93% REPEATS
    Signal

    Most measured interruptions repeated a warning the system had already made.

    Root cause

    The gatekeeper behaved like a latch: once triggered, it kept speaking.

    Change

    I rebuilt it as a circuit breaker that intervenes only for material risk.

What I built

I built one control system for all of my AI agents and projects. It keeps project status current, decides what should happen next, challenges risky plans before work starts, and makes irreversible actions wait for me.

Complete workflow

a project roster snapshot

Agents write project state backevery project
Sippaactive
Checkout Flow Mobileactive
Agent Dashboardactive
Personal siteactive
Merchant Accessdone
Sippa KOL outreachpaused
Telegram group builderarchived
11active
5done
5paused
5archived
2blocked

Each agent updates shared project state as a by-product of its work.

The roster is the live one, counted by state. The post-mortem figures are from the repository's own README.

System design

Critical controls

  1. 01

    State is a by-product of work

    Agent AAgent BAgent C
    28project states update themselves

    If an agent can observe a fact, I do not ask a person to maintain it. The fields that stayed complete were the ones every agent wrote back as part of doing the work.

  2. 02

    Least privilege is part of the product

    RepositoryOne writer
    SendDeletePublishSpend
    Human approval

    Each agent receives only the tools its task needs, only one writer can touch a repository at a time, and publishing, deleting, sending or spending always stops for me.

  3. 03

    A useful gate should be quiet

    93%original interrupts repeated
    Quiet by default
    Material risk only

    The first gatekeeper repeated 93% of its measured interruptions. I changed the design from constant warning to selective intervention so that a warning still carries weight when it appears.

In operation

Operating results

  • 2months from nothing to daily useNine in ten commits are hers
  • 13morning-brief runs turned into a product ruleLed to zero-click delivery
  • 93%repeat interruptions identifiedThe gate was rebuilt as a quiet circuit breaker

The roster is the live one, counted by state. The post-mortem figures are from the repository's own README.