Getting started — your plain-language guide

Welcome. This product lets you run AI agents with confidence — knowing exactly what each one can do, how much it can spend, and that everything is recorded and verified. You don't need to be technical. If you can use a phone, you can use this.

This guide walks you through the handful of things you'll actually do. Each takes a minute or two. Do them in order the first time; after that, come back to any section as a reminder.

The one idea to hold onto: a new agent can do nothing until you allow it. You're never racing to lock things down — you start fully safe and open up only what you choose. Every screen is built around that.

What you'll learn here

1. Add your first agent
2. Allow it to do one thing — safely
3. Set a budget
4. Approve (or deny) a request
5. Stop an agent
6. Get a proof report

There are only six things to know about in the whole product: Agents, Permissions, Budgets, Approvals, Stop, and Proof. That's it. For a quick definition of each, see the six concepts.

A 30-second tour

When you sign in, you land on the Dashboard. It answers one question: is everything okay?

• A big, calm headline tells you the state: "Everything is governed" in green when all is well, or something like "2 approvals are waiting for you" in amber when you're needed.
• Below it: how many agents are Safe, how many Needs attention, how many are Stopped, and total spend this month.
• At the very bottom, a green strip reads "Verified — every action is governed and recorded." (On a phone it shows up top as a small "Verified" pill.) That's your ambient reassurance — tap it any time to see the proof.

Colors mean the same thing everywhere:

📷 Screenshot: the Dashboard — see the prototype Dashboard screen. (The prototype shows the amber "needs you" state — "2 approvals are waiting for you" — so you can see what attention looks like.)

1. Add your first agent

Where: the Agents screen → Add agent.

1. Open Agents from the left menu. If you're new, you'll see: "No agents yet. Add your first — it'll start completely locked down, and you decide what it can do."
2. Click Add agent.
3. Fill in three short fields:
   • Name — what this agent is, in plain words (e.g. "Refund assistant").
   • Owner — the person accountable for it. Pick a real human.
   • Risk level — Low, Medium, or High. There's a one-line explanation under each. (High means it can take actions that are hard to undo. The riskiest actions ask for your approval automatically based on what they do; having your chosen risk level shape those defaults is coming.)
4. Click Create agent.

You'll see a reassuring confirmation: "[Agent] created. It can't do anything yet — you choose what to allow." The app then takes you straight to Permissions, with everything switched off.

That's the point. Your new agent exists, it's named, it has an owner — and it is completely safe, because it can't do a single thing until you say so.

💡 Tip: the Risk level doesn't grant anything — and today it doesn't change behavior yet. The riskiest actions pause for your approval automatically either way; risk-level-driven defaults are coming. You're always in control.

📷 Screenshot: the Add agent form — see the prototype Agents screen.

2. Allow it to do one thing — safely

Where: open your agent (click its row in Agents) → Manage permissions.

Right now your agent can't do anything. Let's allow it one capability — only what it actually needs.

1. From the Agents list, click your agent to open its detail page.
2. Under What it's allowed to do, you'll see "Cannot do anything yet." Click Manage permissions.
3. At the top, a banner reassures you: "Everything is off unless you turn it on. An agent with nothing on is completely safe."
4. You'll see a plain list of capabilities, each switched Off — for example "Read customer data," "Send emails," "Issue refunds."
5. Turn on just the one you want. When you do, a short note appears inline:
   • how risky it is, in plain words ("Refunds move real money and can't be undone."),
   • and whether it asks for your approval each time. The riskiest capabilities are set to wait for you automatically. (Choosing this yourself, per capability, is coming; narrow limits like "only refunds under $500" aren't available yet.)
6. That's it. Leave everything else Off.

The golden rule: turn on only what the agent needs, one at a time. Every switch left off is one less thing to worry about.

💡 Why "off" is good: an off switch here isn't a gap you forgot to fill — it's protection working. The fewer things on, the safer the agent.

📷 Screenshot: the Permissions toggles with an inline risk note — see the prototype Permissions screen.

3. Set a budget

Where: Budgets in the left menu (all agents), or Manage budget on an agent's detail page (just that agent).

A budget is how much an agent can spend in a month — like a prepaid card.

1. Open Budgets, or click Manage budget on your agent.
2. Set a monthly cap — a simple amount. The screen phrases it plainly: "Stop spending around $500/month."
3. Watch the usage bar: it fills green, turns amber as it nears the cap, and red when it's reached.
4. Choose what happens when the cap is reached:
   • Just alert me — keep running. (The in-product alert that tells you is coming; for now, the usage bar on this screen is where to look.)
   • Slow it down — keep going, more carefully.
   • Stop it — halt the agent. (Recommended for high-risk agents.)
5. Save. (A manual reset, and showing exactly when the monthly count resets, are coming.)

⚠️ An honest note about budgets. A budget is a smart limit, not a to-the-cent wall. Spending is watched continuously and acted on quickly, but a small overshoot is possible before a hard stop kicks in. So set your cap as "stop around here," not an exact ceiling. The screen reminds you of this too — it's there so you're never surprised.

📷 Screenshot: the Budget cap, usage bar, and the three on-cap choices — see the prototype Budgets screen.

4. Approve (or deny) a request

Where: Approvals in the left menu. A number badge shows how many are waiting; it turns amber when there's anything to do.

When an agent wants to do something risky that you set to need approval, it pauses and waits for you. Nothing happens until you decide.

1. Open Approvals. You'll see one simple card per waiting action:
   • What it wants to do: "Refund assistant wants to issue a $240 refund to Acme Co."
   • Why it needs you: "Refunds can't be undone."
   • Who and when: which agent, how long it's been waiting, and when the request expires.
2. Click Approve to let it proceed, or Deny to stop it.
3. The card shows a short confirmation and drops off the list. The badge count updates.

A few reassurances:

• If you don't decide in time, the request expires safely. The agent does not sneak the action through. Doing nothing is always the safe choice. (A visible log of expired requests is coming.)
• When there's nothing waiting, you'll see the good state: "Nothing waiting. You're all caught up."

💡 The riskiest capabilities always wait for you here — the product sets that automatically. Choosing for yourself which capabilities need approval is coming, in Permissions.

📷 Screenshot: an approval card with Approve / Deny — see the prototype Approvals screen.

5. Stop an agent

Where: the Stop button on any agent's detail page.

If something ever looks wrong — an agent behaving oddly, or you just want to be safe — you can stop it instantly. Think of it as the big red emergency-stop button on a machine.

1. Open the agent and click Stop.
2. A confirmation appears: "Stop [agent] now?" with a plain explanation: "It will stop immediately, lose its access, and be blocked from acting again until you turn it back on. Anything in progress stops at once."
3. Click Stop now to confirm (or Cancel to back out).
4. You'll see: "Stopped. It can't do anything else." The agent's status turns red everywhere it appears.

What this means in plain terms: the agent isn't just paused on one task — it's fully held back across everything it could do.

Bringing it back: restarting a stopped agent is coming. Today, stopping is one-way — a stopped agent stays stopped, and nothing it was doing resumes. So treat Stop as what it is: the emergency brake.

📷 Screenshot: the Stop confirmation and the "Stopped. It can't do anything else." result — it appears as a confirm dialog on the prototype Agent detail screen.

6. Get a proof report

Where: Proof in the left menu, the green trust strip at the bottom, or View proof on any agent.

Proof is your evidence that everything was governed — a report you can hand to a board member, an auditor, or a regulator. It's like a receipt with a tamper-proof seal.

1. Open Proof. At the top is a Verified badge: "Verified. Nothing was tampered with."
2. Below it, a plain summary — one sentence per kind of activity:
   • "1,204 actions — all governed."
   • "38 approvals (32 approved, 6 denied)."
   • "3 agents stopped — each fully blocked from acting."
   • "Nothing in this record was altered."
3. Today the report covers your full record. (Focusing it by time range or by agent isn't available yet — the controls on the screen say so too.)
4. Click Export to produce a file you can share. You'll pick how much detail to include:
   • Summary only (the default — safest, shares the least)
   • Standard report
   • Full evidence
5. A footer confirms: "This report is complete and verifiable. Anyone can independently confirm it wasn't changed."

⚠️ An honest note about proof. Proof shows what was governed and that the record is complete and unaltered — no one quietly changed it after the fact. It doesn't claim every action was correct in the real world; it faithfully records what happened and seals it so it can be trusted. That honest framing is exactly what makes it trustworthy to an auditor.

📷 Screenshot: the Verified badge, summary, and Export options — see the prototype Proof screen.

You're set up

That's the whole product. Day to day, you'll mostly just:

• glance at the Dashboard to confirm everything's green,
• clear any Approvals waiting for you, and
• occasionally adjust a Permission or Budget.

If anything ever feels off, Stop is one click away — and Proof is always there to show everything was governed.

A good first-week routine:

1. Add one agent.
2. Turn on the single capability it needs (leave the rest off).
3. Set its budget and choose what happens at the cap.
4. Check Approvals once a day.
5. At month-end, open Proof and export a Summary — just to see how easy it is.

You started fully safe, and you only opened up what you chose. That's the whole idea.

Need a refresher on the six core ideas? See the six concepts.