Skip to content
Nick Vidal

All work

Personal project · 2026

Dwight: teaching myself how to monitor AI agents

I built Dwight to teach myself how to monitor AI agents, not just run them. You talk to one assistant, a digital twin acting as chief of staff, and it hands work to a small fleet of agents that research, analyze, plan, and write. What I wanted to learn is everything around them: what they're allowed to do, who approves it, what it costs, and a record nobody can quietly edit.

Dwight's operator console, Approvals page: two pending human approval gates, G1 Agent Creation and G2 Agent Activation, each with Approve and Deny buttons.
The approvals queue. Creating and activating an agent each wait for a person. From a fresh demo install with no personal data; the real console stays on my private network.

What I built

Most of what I wanted to learn sits in the parts you can't see in a chat window: who's allowed to do what, what needs a person's yes, what it costs, and how you'd know afterward.

  • A digital twin, Dwight, is the one assistant I talk to. It can read and search my email, check my calendar, draft replies, read my documents, and research the web.
  • Two tiers of agents behind it. Team leads for research, product, marketing, and operations each break work down for specialist worker agents, and the results come back up.
  • Delegation is guarded: Dwight proposes, the operator confirms, then it dispatches. Nothing consequential starts on its own.
  • Twelve approval gates (G1 to G12) cover creating or activating an agent, granting a tool, installing a skill, writing sensitive memory, relaxing a constraint, any outside action, going over budget, resuming after an incident, spend requests, and host changes.
  • Every action is written to an audit log where each row carries the hash of the one before it, so a change after the fact would show.
  • Token and cost budgets, and going over one is itself a gate that needs approval.
  • Outside systems connect through an MCP gateway. Today those connections are read-only, such as my deploy platform. Server and GitHub access is built behind two gates and not granted to any agent.
  • It's built in phases, with a status file that says what work is allowed right now and what stays forbidden until the operator unlocks the next phase.

Worth knowing

It's a learning project and still being built: it's partway through its tenth phase, and the console stays on my private network. There's no public demo. The real console shows my email and documents, so these screenshots come from a fresh install with an empty database and nothing connected.

I built it with AI coding tools. I wrote the rules it runs under, decide when each phase opens, and I'm the one accountable for what it does.

Dwight's fleet view: the Dwight digital twin card marked proposed, with approval required, and a tile to propose a new agent through a G1 approval request.
The fleet view on the same demo install. Proposing a new agent files a G1 approval request; nothing joins the fleet without a yes.