Employee Zero - The Difference Between Governed and Merely Useful
I love Grokbot from xAI and Dot from OpenAI, I find them very useful. I use them both daily to manage my emails. Grokbot automatically applies to jobs I’m sent that match my criteria and then archives the gazillion job emails. Dot has informed me of money I’m spending on a subscription I don’t use. And much more. I’ve spent a year developing my own AI Employee Factory and pondered - what’s the difference? Is there a reason for custom agents in an era of these powerful general agents from the trillion dollar companies themselves?
Some of you know of my development environment stability issues when developing with MANY simultaneous Codex/Claude coding sessions. I had a PC that wasn’t up to the task and moved my work to my Windows 11 machine using WSL2 Ubuntu Linux. I gave 20gig of the machines 32gig to the WSL2. For a couple of months this machine has put up with hard-core, multi-project development. But yesterday I pushed past its limits. I was a day away from the weekly OpenAI subscription reset and had a lot of capacity to use or lose. My linux session locked up and I had to restart, destroying a bunch of running sessions. Fortunately nothing happened to the system and everything came back up without damage. But WHAT WAS I WORKING ON? What did I need to restart?
I directed my Chief of Staff AI to build me a report of all the work I had made progress on over the last 2 weeks, next step, and a revival prompt. As this is a useful and repeatable need, I directed Chief to create a new AI employee. (I call my AI Agents “employees” because that’s how I manage them, like having employees).
When building my AI Employee Factory I started with the concept of a base or foundational employee that has all the features that ANY employee would need. In object orientation, this would be a class, and new objects would inherit from the class. One of the biggest differences between using GrokBot, Dot, OpenClaw or Hermes and my own agents is this foundational employee template that I call: Employee Zero.
Employee Zero - Base Features
Every employee starts with a common foundation:
- Identity and purpose: Knows its role, responsibilities, and who it reports to.
- Memory and continuity: Preserves context, decisions, and work state across sessions.
- Knowledge access: Consults authoritative sources and distinguishes facts from assumptions.
- Tools and permissions: Uses capabilities explicitly authorized for its job.
- Task accountability: Works toward defined outcomes with clear completion criteria.
- Communication and coordination: Shares progress, exchanges handoffs, and escalates blockers.
- Quality control: Verifies results and submits material work for independent review.
- Bounded execution: Respects resource limits, handles recoverable failures, and stops when it cannot safely proceed.
Specialized employees add their job-specific skills and authority to this foundation.
There is specific code that I’ve matured over time to accomplish those high level tasks. Behind those features is a combination of code, shared services, and operating instructions that I’ve matured over time. One is Agent Board—a shared task and status board where employees coordinate work, report progress, and communicate with the Chief of Staff and me.
Memory includes Markdown files that preserve decisions, current work state, and next steps. If a computer crashes or a subscription runs out, another supported model can use that saved context to pick up the work. A shared, searchable knowledge base also gives employees access to indexed information across my projects, within their permissions.
BTW, I looked into the crash which was related to exhausting the memory and cache to find the culprit. It’s not my Docker containers nor the Python apps the coding harnesses call. It’s Codex, Claude themselves when I’m running a lot of simultaneous projects and the projects are spawning off multiple sub-agents. I’ll create some kind of monitor with alerts to forestall this in the future.
The Working On report that was just created - I’m so glad to have it. I can quickly find what I was working on, the next steps and a prompt to pick back up and go forward.
The general bots will improve over time - and so will my custom “employees”. For personal “do for me” tasks, I believe these personal agents will rule. For work? For companies? Governance and monitoring and quality control will demand the custom work. IMHO.
ps: yes, this is also an example of "AI as vending machine" and "skillfully using AI"
