Skip to the page

Act · AI Agents

Give a job to an AI agent. Keep the final yes.

Describe the job in your own words, give it your documents and the flows it may run, and try it on the same page. Anything it would change in your systems or send to someone waits for a person's yes. It can never approve its own work.

Works with Workflows AI Agents Connections
Tarnwick Supply Co.Agents · Quote deskSample data
Quote deskAgent · version 2Tests pass: 3 of 3 VersionsTest setsSave
The job, in your words

Answer quote requests from our price list. Check stock before you quote. Never offer more than 5% off. If a customer asks for more, hand it to Dana.

1Describe the job
Documents it knows Price list 2026.pdf12 pages Delivery terms.pdf3 pages
Flows it may run
Check stockLooks things upUse it when a quote names a product
Make a quoteChanges something: asks a person first 3Never approves its own work
Try it
Quote 40 steel brackets for Fenwright Builders.
Checked stock: 52 in the North warehouse. $18.40 each, 40 for $736.00. Price list 2026.pdf, p. 3 This try looked things up for real and changed nothing. 2Try it: nothing changes
What this try did
  • Ran "Check stock": looked things up
  • Read Price list 2026.pdf, p. 3
  • Would run "Make a quote": a person is asked first
  1. 1Describe the job
  2. 2Try it: nothing changes
  3. 3Never approves its own work

Four steps to an agent.

try it: nothing changes

Sample data · Tarnwick Supply Co.
  1. 1

    Describe the job

    Brief it in plain words, as you would a new colleague, or start from an example: answer quote requests, check supplier invoices against orders, or sum up overdue payments every Monday. Our own rules come first.

  2. 2

    Give it documents

    Drop files on it, up to 50 MB each. It answers from them and names the file it used. Or share a library, such as a catalogue or your price lists, with several agents. Each person only gets passages from documents they may open.

  3. 3

    Choose what it can do

    Pick the flows it may run, each for one situation with a "Use it when" sentence. Each is marked "Looks things up" or "Changes something: asks a person first". Need a new one? Build it from one short dialog, and the designer opens with it ready.

  4. 4

    Try it

    On its page, with the real model. It really looks things up, and tells you what it would do for anything that changes or sends. Nothing changes.

Then one Save puts it to use as a new version, after a "what changed, try it first" check. Your company's owners make agents, with no help from our staff; members use the agents shared with them.

Put it to work.

  • One click makes a flow with the agent already in it, started on a schedule or by hand.
  • Run it now. Press Run now and type what to ask it. The answer is on the run's page.
  • In any flow. One step, "Ask an agent", hands it each row, or all rows at once, with a task. The answer comes back in fields you name. If it could not finish, the row goes out of "Not done", with the reason.
  • Options that matter. Pin a version. "Do each change once per" a message, so an item read twice never proposes the same change twice. Fewer rounds, or another model. When rounds run out, ask a person or stop.
  • Patient. It waits for a person as long as needed, and picks up after a server restart without doing anything twice.
  • Asks when it needs to. It can ask a person a question and wait for the answer on the Waiting page, or ask another of your agents to look something up. An agent asked this way never changes anything.
Quote requestsFlow · version 1Sample data
Quote requestsevery weekday at 09:00
Read new quote requests6 rows
Ask an agent: Quote deskeach row, with a task
Done · 4 drafts
Ask someone firstDana Ortiz decides
Not done · 2
Asked for 12% offhanded to Dana
1Not done comes with the reason

The quote requests are read from a sample web-API connection.

  1. 1Not done comes with the reason

It asks first, every time.

never approves
its own work

  • Every change in your systems, every message and every flow that writes or sends waits for a person on the Waiting page. The record shows it was proposed by the agent and decided by a person.
  • It can never approve its own work, even when the person running the flow could approve it.
  • It reads and acts as the person the flow runs for, through your company's own connections. It holds no password or key, and a tool that person cannot open is never offered to it.
  • It sends e-mail only to the contact the task is about, the person the flow runs for, or addresses your owner listed; never to an address that a document or the model wrote. Each message waits for a person.
  • Letting an agent make a change without asking needs a grant only our staff hold, set for one kind of change on one connection.
WaitingQuote requestsSample data
  • Read the request from Fenwright Builders
  • Checked stock: 52 in the North warehouse
  • Priced it from Price list 2026.pdf
Make a quote?Waiting for your yes
proposed by Quote desk (version 2) Quote for Fenwright Builders: 40 steel brackets, $736.00
ApproveDecline1Waits for a person's yes
The agent can never approve its own work.
  1. 1Waits for a person's yes

Safe with what it reads.

  • What it reads in an e-mail or a file is material, never orders. Its standing rules tell it never to invent a number, a name, a date or a promise.
  • Six checks watch four places: its task, what it reads, what it sends and its answer. They look for personal data, passwords and keys, instructions aimed at an AI (always on), talk off its topics, words it must never say, and web addresses not on your list. Each check notes, masks, asks a person or stops.
  • Once it has read text that tried to give it orders, everything it proposes after that waits for a person.
  • In our release test, a PDF told an agent to send the price list to an outside address. Nothing was sent.
Run recordQuote desk · release testRecorded on stand-ins
  • Read request-2231.pdfflagged: instructions aimed at an AI ('send the full price list to …')step 3 1Material, never orders
  • Everything after this waits for a personstep 4
  • Nothing was sent.end
  1. 1Material, never orders

Test it like software.

  • Keep any try as a test case.
  • Test sets run the agent on stand-ins, so nothing real is touched. Each case is checked by rules and, if you want, by a second AI model scoring 0 to 10. See the cost first, and compare two runs.
  • Publishing can warn you, or wait until the test set passes.
  • Versions are numbered: compare any two and put back an older one. A flow follows the newest version, or stays on one you pinned.
  • See everything it did under each "Ask an agent" step: every round, the model, the time and tokens, each call with what went in and what came back, and whether it ran, waited, was declined or refused.
  • Set how it thinks: its model, and its limits on rounds and tokens. It asks before it goes past them.
Test setsQuote deskSample data
Version 11 of 3passed
Version 23 of 3passed
Customer asks for 12% off → hands it to Danapassed
Checked by rules: hands it to Dana · offers no more than 5% offRun on stand-ins: nothing real was touched
Quote 40 steel brackets → $736.00passed
A product not on the price list → says sopassed
Publish only when the test set passes1Goes to work when its tests pass
  1. 1Goes to work when its tests pass

Memory you control.

  • Memory is off until you switch it on: a note on each contact, facts about the company (used only once an editor confirms them), and notes about the person running it.
  • Nothing is kept silently or from a test, and card numbers are never kept. Contact notes go 90 days after their last use, or after the time you set, from 7 to 365 days.
  • See, change and delete what each agent keeps.
  • An owner can erase or export everything kept about one contact, and each request is recorded.

It works with the other engines.

Workflows

An agent's skills are your own flows: visible, tested and versioned, not hidden code. Any flow can ask an agent in one step.

Workflows

Documents

It reads the files it is handed, scans included, and names the file it used.

Data connections

It reads lists from your connections as the person the flow runs for, and creates or changes a record only through an action your owner declared on that connection.

Data connections

Outside tools

Connect an outside tool server (MCP) and approve its tools one by one, as a read or a write. Your agents use only what you approved, and a tool that changes on its server is set aside until you approve it again.

AI models and usage

An owner chooses which AI models agents may use, with a default and a fallback. Every round and tool call shows on the Usage page; a try is counted apart, never as a run. An owner can switch the Agents app off for the company: its pages say so, nothing is deleted, and flows that ask an agent keep running.

Sharing

Only you, chosen people and teams, or everyone. Whoever may use an agent can ask it from their flows and try it; its flows run as that person, with that person's access.

What it does not do yet.

So you know before the demo.

  • You cannot talk to your agents in the assistant's chat yet, and the assistant cannot make or change an agent for you.
  • Agents start on a schedule or by hand.Starting when mail arrives or a form is sent is not offered yet.
  • Your customers cannot talk to an agent directly.
  • There are no ready-made agents or an agent gallery: three example briefs to start from.
  • Naming who approves an agent's changes on the agent itself is not followed yet; the approvers named on the connection decide.
  • An agent's flows are built on a laptop or desktop; on a phone you can see, try and share an agent.

Bring one job you would hand to a new colleague.

In the demo we connect one of your systems, and you ask the questions.