6 steps, of which 4 are a rule rather than a judgment. Rebuilt as 2 person/people and 3 agents, with one named human accountable for every agent.
The process
The steps
What happens now, and what happens after
The question that matters
For each step, one question. Does this need a human judgment, or is it a rule being
followed? If it is a rule, a machine can do it, and it will do it the same way every
time. If it needs judgment, a person keeps it. Where a step is a rule with a judgment
sitting inside it, an agent prepares the work and a person decides.
Before
After
A step with a red edge on its right cannot be undone. Those keep a
named person on them whatever their nature, because liability cannot be delegated.
The map
Who is on this, carbon and silicon
What an agent is
An agent is a piece of software that has been given a job, some tools, and permission
to use them. It is not a chatbot you talk to. It is closer to a new hire who has been
told what their job is, what systems they can log into, what they are allowed to
change, and who to go to when they are stuck. Every agent below has exactly one named
human accountable for it. You can delegate the work. You cannot delegate the
consequence.
A person An agent Accountable for Reaches a system
The agents
What each agent's job actually is
The agent role definition
This is the part most people skip and it is the part that matters. Before you build an
agent you write down what it is for, what it must never do, what systems it can touch,
the limits it works inside, who it escalates to, what it remembers, how you would know
it is working, and what it writes down about its own actions. If that sounds like a job
description, that is because it is one. You would not hire someone without writing down
what the job is.
Permissions
Who may touch what, and why
The four levels
Read means it can look. Draft means it can prepare something a person sends. Write
means it can change data. Commit means it can do something you cannot take back. Almost
everything should be read or draft, and every commit needs a named human who approved
it. Every line below carries the reason it was granted. A permission with no reason
gets removed.
Agent
System
Level
Why it has this
The people
What the humans on this do now
The run
One agent, run against the spreadsheets
What you are looking at
This agent ran against spreadsheets standing in for the real systems, not against the
real systems. Its reasoning step is a scripted stand-in rather than a live model call.
Everything else in it is real. The permissions were enforced, the guardrails held, and
an escalation would have gone to the named human. Green is a row it created, yellow is
something it changed, red is something it deleted.