Operations Design for Multiple AI Projects

Observation LogOBSERVATION RECORD / 922D38D2RECORDED : 2026-07-20DOMAIN : IMPLEMENTATIONSTATUS : ARCHIVED

Multi-AI development only works when you
Automate Progress Tracking and Approval Managementhave a system for tracking progress and managing approvals.

When multiple AI and Codex projects run at the same time, progress management reaches its limits before work speed does. Which projects are moving? What is blocked? Where is my judgment needed? A process where a person has to enter this information every time will not last. So I built an AI Control Room that automatically gathers progress, execution evidence, pending approvals, project records, and lessons from failures across projects.

AI Control Room project board showing multiple AI projects grouped by status — Projects are grouped as not started, in progress, on hold, stopped, or complete; pending approvals and unfinished tasks appear on the same screen.

Section 01 — A Control Room That Does Not Depend on Manual Updates

The Most Important Part Is Not the Dashboard, but a System Where AI Records Work Automatically

This control room is not a board where people register each project and update its status one by one. When AI starts work, a start event is recorded. When implementation or verification finishes, the outcome and evidence are added. When a decision is needed, an approval request is raised.

People review only exceptions and approvals that AI cannot decide. Once a person approves or rejects a request, the decision is sent back to the project and work within the permitted scope can resume. The design leaves decision boundaries to people instead of unconditionally automating risky external actions or irreversible operations.

If people still have to enter progress reports manually after delegating the work to AI, the bottleneck remains. The first thing to automate is therefore not only the work itself, but also the creation of information used to manage that work.

In multi-AI development, automating tasks alone is not enough.
Progress management itself needs to be automated.

Section 02 — What Happens During Parallel Development

As Projects Multiply, Managing Them from Memory Quickly Breaks Down

With one project, I can reread the conversation and recall its status. But when several AIs work on separate projects at the same time, I cannot manage them in my head alone.

Manual Management

Copy progress updates for each project
Search the history to find what each AI did
Find pending approvals in conversations
Failure reasons are scattered across different places
Repeat the same decisions over and over

AI Control Room
Automatically record start and completion events
Link deliverables and verification evidence to each project
Escalate only exceptions that require a decision
Record failures and remedies in the project record
Carry past decision criteria into the next task

Section 03 — Project Board

Show Not Only Project Status but Also Why Work Is Blocked

Columns for not started, in progress, on hold, stopped, and complete are not enough. Each card should show the next task, unfinished work, whether approval is pending, and evidence of deliverables. If a project is stopped, the card should also record what it is waiting for.

With this setup, I can tell which project needs attention as soon as I open the board. Even if many projects are in progress, there is no need to interrupt if none needs a human decision. Conversely, if even one approval is pending, I can focus on that decision first.

Section 04 — Approval Gates

Separate What AI Can Handle from What Requires a Human Stop

Step 01

AI moves the work forwardAI handles safe local work, research, implementation, verification, and record keeping.

Step 02

Escalate issues that need a decisionOnly actions that require human responsibility—such as publication, external communication, deletion, or major policy changes—are sent for approval.

Step 03

A person approves or rejectsReview the goal, risk, evidence, and recommendation; make a decision and return it to the project.

Step 04

Resume work within the approved scopeCarry the approval decision forward and continue permitted internal work. Any final external action must stay within the established permission boundary.

Section 05 — Project Records

Bring Together What Happened, What Was Learned, and What to Do Next

A project record brings together the goal, constraints, current status, deliverables, execution evidence, unfinished tasks, approval history, decision criteria, and lessons learned. It is more than a task list: it gives another AI enough context to resume the work later.

A report that says only “the work is complete” is not enough to judge quality. Recording what was checked, what evidence exists, and what uncertainty remains helps reduce the risk of work that is merely assumed to be finished.

AI Control Room project record summarizing goals, tasks, evidence, approvals, and lessons — Each project record consolidates its goal, roles, next decision, evidence, approval history, and lessons learned.

Section 06 — A System That Prevents Repeated Mistakes

If Failures Are Not Recorded, AI Will Repeat Them in Another Project

AI can work quickly, but without the right assumptions and constraints, it may repeat an earlier mistake somewhere else. If a person has to say “I already told you this” every time, the benefits of parallel development disappear.

So I record not only what went wrong, but why it happened, what to check first next time, and which actions require approval. In the next project, I give AI those lessons as pre-work cautions and completion criteria.

The point of systematically collecting lessons in one place is not to create more records. It is to reduce repeated decisions and failures, and make the overall operation stronger as more AI work is done.

When several AI agents are working at the same time,
people should not spend their time entering progress updates.
A Person Decides on ExceptionsThey should focus on deciding what to do when exceptions arise.