WEALTH EXPERIMENT / A PRACTICAL FIELD GUIDE
How an AI agent
gets work done.
You set the goal. The system works in steps. You decide what is ready.
YOUR BRIEF
Result + context + allowed actions + approval points + definition of done
THE HARNESS
Software that runs the work around the model.
Model (LLM)
Generates an output
or a tool request.
Context
Brief, files, history,
and retrieved sources.
Tools
Browser, app actions,
search, and code.
Optional subagents: delegate a smaller task and collect its result.
Choose the
next step
Check access
and approvals
Run the
allowed tool
Observe
the result
Repeat with the new information
Stop when finished, blocked, or at a time, step, or cost limit.
Good controls: permissions, progress tracking, logs, checks, and stop rules.
YOUR REVIEW + APPROVAL
Check the evidence. Approve before publishing, sending, spending,
or changing important data. Verify the final result.
OPTIONAL SUPPORT / ADD WHAT THE TASK NEEDS
RAG
Retrieve sources.
Add them to context.
Check accuracy.
Skills
Reusable instructions
and resources for
a type of task.
Plugins
Packages that may
bundle skills, tools,
and connections.
MCP
A standard for an AI
app to connect to
tools and data.
INFERENCE
One run of the trained model
on the input it receives.
TOKENS
Units the model processes. More context
and repeated runs can add cost.
START SMALL: one task, one tool, one clear review point.
