The AI, built in
Agents that read your procedures and do the work
What it is
Built from your procedures, and it says so
- It follows your procedure, not its instincts
- An agent is compiled from the Documentation your team approved — the decisions and the reasoning behind each step, not just the clicks. The steps it takes are the steps you wrote.
- It shows you which procedure it used
- Answers cite the Document behind them, and every run leaves a record of what it did and why.
- It stops rather than guesses
- When a job falls outside the procedure, or hits a step you marked for approval, it parks and waits for a person.
What an agent can actually do
Opens and writes real documents
Word, Excel, PowerPoint and PDF. It edits the actual file, so your formulas, your formatting and the sheets it never touched all survive. It can fill in a PDF form, and put your fonts, colours and logo on what it writes.
Does the sums properly
It evaluates genuine Excel formulas across every row and reports the result: SUMIFS, XLOOKUP, INDEX/MATCH. Ask a chatbot to total a column and it adds up the rows it happened to read, then calls that the total.
Answers “how do we do this here?”
Agents can search and read your own library at the moment a question calls for it. The procedures you already wrote become answerable, without anyone pasting them into a prompt.
Sends it, files it, posts it
Message Slack, send email, raise a Jira issue, open a Zendesk ticket, update a record, or call any HTTP API you point it at. It finishes the task itself, instead of leaving you a summary to act on.
Drives software with no API at all
On the desktop it gets windows, keyboard, mouse, the screen and the file system. Most business software has no API, so most automation tools cannot reach it — this one can, because it uses the screen instead.
Draws what you asked for
Ask for a diagram, a chart or an image and you get a picture — not a block of source code that renders into one somewhere else.
You control whether each one runs on its own or asks first
Tools are enabled per agent, and any of them can be set to ask first. On the desktop that is a card in front of the action. In the cloud the run stops and waits for a person rather than guessing. Each agent also carries its own step ceiling, per-run spend limit and daily run cap, so a mistake stays cheap.
It starts without you
What can start a run without you
A schedule
Pick a time and it runs then — 6am on the 1st, every weekday at five, hourly. Nobody has to remember it is month end.
An email arriving
Point an agent at a mailbox and it starts when something lands in it. The request, the attachment and the sender are the input.
A form being submitted
A hosted form anyone can fill in — no login, no credential, nothing for the person submitting it to install.
A file appearing in a folder
A watched folder, so dropping a statement or an invoice in is the whole of the trigger.
A record changing elsewhere
Steptree polls the systems you have connected and starts the agent when something it cares about moves. No per-vendor wiring on your side.
A webhook, a calendar event, or a person asking
Anything that can POST can start a run. So can a meeting starting, a message in Slack or Teams, or you typing in the chat window.
You stay in control
Nothing happens that you did not allow
- Tools are granted one by one
- An agent holds exactly the tools you switch on. A tool needing a credential you have not supplied stays off until you supply it. A switched-off tool says so plainly, rather than leaving the agent to insist the capability does not exist.
- Any tool can be set to ask first
- On the desktop that is an approval card in front of the action, describing exactly what is about to happen. In the cloud the run parks and waits for a person — and it can ask a free-form question, not just offer buttons.
- A step ceiling, a spend cap and a daily run cap
- Each deployment carries all three. If an agent fails, the cost stays capped, and the ceiling is enforced where the spending happens rather than trusted to the runner.
- It cannot publish itself to the cloud when it needs a screen
- Steptree reads the procedure first. If a step genuinely needs a desktop, the cloud deployment is refused. The refusal names the skill and quotes the step, rather than shipping something that could never have worked.
- Every run is on the record
- What started it, what it did, which procedure it followed, what it cost, and how it ended. Runs are visible to the team the agent belongs to, so you can see what it did overnight instead of piecing it together afterward.
- It is judged on whether it got there
- A run can state the objective it is working to and the checks it should be judged against, settled against evidence Steptree actually holds. “We could not tell” is a distinct answer from “it failed”, because those two lead a reader to different actions.
Nothing deploys itself
Questions
About agents
- How is this different from pointing ChatGPT at our documents?
- A general model answers from what it has read, and it will answer confidently when it has read nothing relevant. A Steptree agent is built from procedures your team wrote and signed off. It follows your steps and names the one it used. When a job falls outside that procedure, it stops and asks a person instead of inventing the rest.
- Will an agent do things without asking me?
- Only what you allow. Tools are granted individually and any of them can be set to stop and ask first. Nothing runs at all until you deploy it, and every deployment carries a step ceiling, a per-run spend limit and a daily run cap.
- Which AI model does an agent use?
- Whichever you pick, from over two hundred across more than twenty labs. You can change it whenever you like, because the agent is built from your procedure rather than tuned to one vendor's model. Every plan reaches every model.
- What happens when an agent gets something wrong?
- You see it, because every run records what it did and how it ended. Then you fix the procedure rather than the agent: correct the Document, and the next run follows the corrected steps. A procedure can be corrected directly; a hidden prompt usually can't.
- Can an agent work in a system with no API?
- Yes, on the desktop runtime — it gets a real keyboard, mouse and screen on your own machine, with an approval card before each action. Most everyday business software has no API, and this reaches it anyway.
- Can my team just ask it a question?
- Yes — in Slack, Teams, email, a chat widget on your own site, or in the app. It answers from your approved procedures and names the one it used, so anybody can check the answer against the document.
Next
Where to go from here
Cloud or desktop
The one real axis: nobody present and no screen, or your own machine with a person watching.
What they can reach
Channels, connected systems, and the applications that have no API at all.
No model lock-in
Over two hundred models, every plan reaching all of them, switchable per agent.
Approve it once. Let it run on its own schedule.
Record one job free and see the agent it builds follow your exact procedure, tools and limits. Paying is for when it is running on a schedule, watching a mailbox or answering in Slack — for every job like it, not just the one you tested.
