Steps an agent can use

Every step you can add sits in the Nodes panel on the left of the builder, grouped by what it does. Each group heading shows how many steps it holds. Drag a step onto the canvas or click it to add it, then select it to set it up in Details. Finding your way round the builder itself is covered in Find your way round the canvas.

app.centrus.ai/app/agents/workflow-builder

Agents

Build modular AI workflows with reusable nodes

Weekly report for Northgate Housing

1,240 / 10,000 creditsValidateResetDraft savedRunsSettingsSave and Publish

1

Monday mornings

Weekly at 07:00 · Mon

Fetch last week's jobs

List jobs

Completed jobs only

Keep ALL of 2 rule(s)

Write the summary

GPT-5.6 Terra

Build the report

Northgate-weekly-report.pdf · A4

Email the report

1 action enabled

AI builderMap

Ready to publish

2

3

1. 1The Nodes panel. Steps are grouped as Trigger, Knowledge, AI, General, Integration and Control.
2. 2The agent on the canvas. Each card is a step, and the lines set the order they run in.
3. 3Details, showing the settings of the selected step.

• Knowledge steps search the files you have trained Centrus on, or the live web.
• AI steps hold instructions for the model: Intelligence reads, decides and writes, and AI Extract Collection turns free text into a tidy list. Voice AI is the voice on a phone line and is covered in Make outbound calls and answer incoming calls.
• General steps read and make files, run a task in a web browser, look up postcodes and reshape, check and filter data. Send SMS and Outbound call also live here.
• Integration steps read from and write to your connected systems, from Outlook to BigChange.
• Control steps add conditions, loops, approvals and error handling. See Conditions, loops and approvals.
• Trigger steps decide how an agent starts. See Choose how an agent starts.

The figures follow one agent, Weekly report for Northgate Housing. Every Monday at 07:00 it fetches last week's jobs for Northgate Housing from BigChange, keeps the completed ones, has the AI write a short summary in the house style, turns that into a PDF and emails it to Dylan Hughes, the account manager. BigChange stays the system of record: the agent only reads the jobs.

Data moves along the lines between steps as a list of items. Each item is one record, such as one job, one email or one file, with its fields and any file attached to it. A step receives the items from the step before it, does its work and passes items on.

Fetch last week's jobs

List jobs

Completed jobs only

Keep ALL of 2 rule(s)

Write the summary

GPT-5.6 Terra

23 items19 items

• A step that lists records, such as BigChange List jobs, passes on one item for each record it finds. Here 23 jobs come out as 23 items.
• Most steps then work through every item in one go. Filter keeps or discards each one, File Parser reads each file, and Microsoft Excel's Append rows writes one row for each item.
• An AI step reads all the items it receives together, as data, and writes one answer. Here the AI sees all 19 completed jobs and writes one summary.
• Steps that send, create or update something once, such as Send email or Create PDF, act on the first item that reaches them. To act once per item, for example to email each customer, put the step inside a Loop. See Conditions, loops and approvals.

To use a value from an earlier step, open Step data references in the later step's Details and click the field you want. Its reference is copied, ready to paste into a setting such as an email subject or a file name. A reference looks like {{ steps.integration-bigchange_rest-1791540900000-2.json.reference }}. Inside a step you can also refer to the item it is working on as {{ item.json.reference }}, and to whatever started the run as {{ trigger.json.triggeredAt }}.

The fields listed under each step are the ones Centrus knows it usually produces. A connected system often returns more. Run the agent once with Run now in the Runs panel and open the run to see the real field names.

File Search searches the files you have trained Centrus on and returns the most relevant passages. You choose which knowledge it can reach by tag: under Knowledge tags, pick one or more tags. Admins see every tag in the organisation; everyone else sees only the tags they have access to. Training files and tagging them are covered in Upload files and Tags and who sees what.

Web Search searches the live web for up-to-date information. There is nothing to set up.

Both steps work best attached to an AI step's Tools handle, so the AI decides when to search and what to search for, as described under Give an AI step tools. Placed on the main path instead, a search step runs one search with a fixed, general request, using the items it receives as context, and passes on its findings, the query it ran and its sources.

• Example: Report style guide is a File Search step tagged Report templates and Northgate Housing. The AI step calls it to check how Northgate Housing likes its weekly reports laid out.
• Cost: measured by the work done, at least 2 credits for each search, whether it runs as a step or the AI calls it as a tool.
• A File Search step with no tags selected gets the validation warning File search is enabled but no tags are selected.

Intelligence is the step that thinks. It reads the items it receives, follows your instructions and decides what to write, extract or reply. It can call tools along the way, such as searching your knowledge or looking up a job. Most agents have at least one.

• Model. New AI steps use GPT-5.6 Terra, a good balance of quality and cost for most work. GPT-5.6 Sol and GPT-6 Astra are stronger and cost more; a lower-cost GPT-5.6 model suits simple, high-volume steps. How the models compare is covered in Choose the AI model.
• Reasoning level. For GPT-5.6 models the choices run from None (Fastest), the default, through Low, Medium, High and Extra High to Max (Most thorough). GPT-6 Astra starts at Low (Recommended). Higher levels take longer and use more credits, so raise it only when the answers need more thought.
• Instructions. Say what the step should do, which tools to use and what the answer should look like. Paste references from Step data references to point at earlier data, such as the invoice reference from BigChange.
• Max tool iterations. 10 by default, between 1 and 30. If the AI reaches the limit, it stops and the item leaves on both the main output and the overflow output, marked as cut short, so you can catch it.
• Preserve incoming item fields. Off by default. Turn it on when later steps still need the original fields, for example the job ID alongside the AI's triage note.
• Emit each item from field (optional). If the AI's answer contains a list, such as jobs, enter that field name and the step passes on one item per entry. Copy parent fields into each emitted item and On empty array then appear.

The AI's answer comes out as content, the text it wrote, and structured, the named fields it returned when your instructions ask for them. Named fields are also copied to the top level of the item, so a later step can use {{ item.json.summary }} directly.

• Example: Write the summary receives the 19 completed jobs and is told to write a 150 word summary for Northgate Housing in British English, grouped by site, using the style guide.
• Cost: measured by the work the model does, at least 2 credits each time the step runs. Each tool it calls is charged at that tool's own rate.

An Intelligence step has a Tools handle on its top edge. Connect File Search, Web Search or an integration step to it and the AI can call that step while it works, as often as it needs, up to Max tool iterations. A tool is not part of the main path: it runs only when the AI asks for it.

Report style guide

2 knowledge tags

Read job notes

List notes

Completed jobs only

Keep ALL of 2 rule(s)

Write the summary

GPT-5.6 Terra

Build the report

Northgate-weekly-report.pdf · A4

ToolsMain outputAI overflow / safety fallback

• An integration step attached as a tool shows Choose the tools this assistant can call. It fills the fields on the call. with a Select tools list. You can offer the AI several actions from one system, and it fills in their fields itself.
• Turn on Require approval and tick Pause for approval to have someone approve each call before it happens: When the AI uses this, pause for someone to approve each time. Approve or reject from the Runs panel. Use it on tools that write, such as creating a BigChange job or sending an email.
• Big lists such as every open ticket or every invoice are not offered as tools. Fetch those with Ingest instead, described under Large lists.

Require approval only takes effect on a step the AI calls as a tool. On a step in the main path it does nothing. To have a person check a write on the main path, put an Approval step in front of it. See Conditions, loops and approvals.

AI Extract Collection reads free text, such as an email body, notes or a web page, and pulls out one or several records as a clean list, one item per record. It has the same Model, Reasoning level and Instructions as Intelligence, but no Tools handle, and its settings start ready for the job: Emit each item from field (optional) is set to items, and Preserve incoming item fields and Copy parent fields into each emitted item are ticked.

List the repairs

AI extract from items

Raise each repair

Create job

3

Ask Sarah for details

1 action enabled

4

Tell the office

1 action enabled

5

1. 1The field in the AI's answer that holds the list. Each entry becomes its own item.
2. 2Each item keeps the fields of the record it came from, such as the sender of the email.
3. 3Extracted items, one per record found.
4. 4No collection found, when the AI finds nothing to extract.
5. 5Invalid source value, when the answer has no list where you said it would be.

• Example: Sarah Price at Riverside House emails four repairs in one message. List the repairs turns them into four items, each with the flat number, the fault and how urgent it is, ready to raise as four jobs.
• Cost: measured by the work the model does, at least 2 credits each time the step runs.
• For text inside a file, use File Parser's structured fields instead of adding AI Extract Collection after it.

Three General steps work with files. They act on files already on the item, for example after Outlook's Download attachments, a download from OneDrive or BigChange, or a file someone uploads in chat.

File Parser reads each incoming file with AI and returns its text, or the specific fields you ask for.

• File types to parse covers Images (JPG, PNG, WEBP, GIF), PDF, Word documents (DOC, DOCX), Spreadsheets (CSV, XLS, XLSX) and Text (TXT, MD, JSON, HTML).
• What to extract says what you want, such as Extract the supplier, receipt date, line items and totals. Leave the structured fields empty to get the full text back instead.
• Structured fields (optional) gives you clean fields to use later. A field called supplier comes back as {{ item.json.supplier }}.
• A file of a type you have not ticked, or a missing file, is not read, but the item still carries on to later steps marked as not parsed. Tick Stop the branch when the file is missing or not a selected type only when nothing after this step should run without a parsed file.
• A spreadsheet already in OneDrive or SharePoint is better read where it is, with Microsoft Excel's Read rows.
• Cost: measured by the work done, at least 2 credits each time the step runs.

Extract Images pulls the pictures out of PDF, Word and PowerPoint files, and passes on images that are already attachments. Engineers' site photos inside a PDF job report are a typical use. Create PDF turns an HTML and CSS template into a finished PDF, such as the weekly report.

• Extract Images: pictures come out one item per picture by default, ready to upload to OneDrive, SharePoint or a BigChange job. Choose Collect: all pictures on one item to keep them together, for example to attach them all to one email. Max pictures is 50 by default and can go up to 100. A file with no pictures leaves on the Empty exit, and you can tick Stop the branch when nothing is extracted instead. Cost: 1 credit each time the step runs.
• Create PDF: set the File name (.pdf is added if you leave it off), Page size (A4, A3, A5, LETTER or LEGAL), Orientation and Margins (points), then the layout under Advanced: HTML & CSS template. The template can use references to fill in the data. Each run of the step makes one PDF, so to make one per item, put it inside a Loop. Cost: 1 credit each time the step runs.
• To email a PDF, tick Attach files from incoming item binaries in the Outlook or Gmail Send email step that follows.

Browser Use has a task carried out in a real web browser, for systems that have no integration: logging in to a customer's contractor portal to fetch a purchase order number, say, or filling in a supplier's web form. You describe the goal in plain words and it works through the pages itself.

• Credentials. Save logins and keys in the step's Credentials section, as a username and password or as a single value, then refer to them in the goal by name, such as %portalUsername%. Never paste a password into the goal itself. A goal that names a credential the step does not have gets a validation warning.
• Goal and System prompt (optional). Write the goal as numbered steps. The system prompt is for standing rules, such as never submitting a form without a reference.
• Timeout (minutes) is 10 by default, up to 30. Maximum steps is 20 by default, up to 100. A session that runs out of time fails with Browser task did not finish within its configured timeout.
• Use proxy is off by default. Turn it on only if a site blocks the browser.
• Structured result fields come back under result, so a field called poNumber is {{ item.json.result.poNumber }}.

Browser time comes from a monthly allowance of 300 browser minutes for your organisation. Every session counts at least one minute, and each started minute counts in full. The bar turns amber at 90% and red at the limit, when New browser sessions are paused until the monthly allowance resets. Extra blocks of 300 minutes are available as an add-on. Each session is also charged in credits. See Automation credits and usage limits.

Geocode Postcode turns a UK postcode into a latitude and longitude, for example to work out which engineer is nearest a site or to put jobs on a map. It does every incoming item in one go.

{
  "reference": "J48213",
  "site": { "name": "Riverside House", "postcode": "HU1 2AB" },
  "geocodeSuccess": true,
  "postcodeInput": "HU1 2AB",
  "postcodeNormalized": "HU1 2AB",
  "outcode": "HU1",
  "latitude": 53.7437,
  "longitude": -0.3324
}

• Source field path is where the postcode is on each item, such as site.postcode. Or enter a Fixed postcode (optional) to use the same postcode for every item.
• Allow outcode centroid fallback uses the centre of the postcode district, such as HU1, when the full postcode is not found.
• Each item gains latitude, longitude and geocodeSuccess. An item whose postcode cannot be found still carries on, with geocodeSuccess false and a geocodeError, so check it with a Filter if later steps depend on it.
• Cost: 1 credit each time the step runs, however many postcodes it looks up.

Five General steps work on data without any AI, so they are quick, predictable and cost 0.1 credits each time they run. Use them in place of asking an AI step to do the same job.

• Filter keeps items that match your rules and sends the rest out of the Discarded items exit. Comparisons include equals, contains, greater than, date before, within last N days, is today and is this month.
• Validate checks items against rules and splits them into Valid items and Invalid items, for example to send records with a missing postcode to someone to fix.

• Extract Collection turns a list inside an item, such as an invoice's lineItems, into separate items. Leave Source field path (optional) blank to use the whole item. It has the same three exits as AI Extract Collection.
• Transform reshapes fields with four operations: Set field, Rename field, Remove field and Keep only fields. Values can be fixed text or references.
• Aggregate brings items back together. Collect all items into a list makes one item holding them all, Count items by field value counts them by a field such as status, and Group rows (reduce) groups by a field and adds up, finds the earliest or latest, or joins text.

A Filter or Validate with no rules lets everything through, and Transform with no operations passes items on unchanged. Validation warns about the first two, but the agent will still run.

Some jobs need thousands of records: every outstanding invoice, or every job in the last 12 months. Passing that many items from step to step would be slow and costly, so Ingest fetches the full list from a connected system in the background and stores it as a temporary dataset. A Loop set to Dataset mode then reads the rows in batches, and the AI only ever sees one batch at a time.

• List source offers the large lists from BigChange, BigChange Web Services, Simpro, JobLogic, Sage Intacct, Xero, Freshdesk, Fixflo, Cin7 Core and GoHighLevel, each with its own filters, such as Date window and Outstanding only for BigChange invoices.
• Ingest passes on a single item with the dataset's datasetId and rowCount, not the rows. Point a Loop in Dataset mode at it to work through them.
• A dataset holds up to 50,000 rows and is kept for 7 days.
• Merge datasets combines staged datasets in one step: Restrict to an id set keeps rows whose field appears in another dataset, Enrich from a lookup copies chosen fields across from a matching row, and Union datasets joins datasets end to end.
• Cost: 0.1 credits each time each step runs. The Loop that reads the batches is free; the steps inside it are charged on every pass.

An integration step reads from or writes to one connected system, such as Outlook, Excel or BigChange. Centrus works alongside these systems: the records stay where they are, and the step only reads or changes what you set it to.

1. 1Save the agent as a draft. The connection belongs to the saved agent, so a new agent cannot connect until then: Save this workflow as a draft first, then connect the account.
2. 2Select the step and click Connect in its Connection section. Microsoft, Google and Xero accounts open their own sign-in page. The others open a form for the API details the system gives you, with Copy from existing agent to reuse details from another agent you created.
3. 3Choose the action under Actions, or the endpoint under Endpoint tool type.
4. 4Fill in the fields the action needs, typing values or pasting references from Step data references.

• Read or write. Actions that list, get, find or download read data and pass it on. Actions that create, update, send, move or delete change something in the other system. A list that finds nothing leaves on the Empty exit (Empty: no matching rows), so you can handle it separately.
• One action per step. On the main path a step runs one action. To do two things in Outlook, such as send an email and then file the original, use two Outlook steps. Attached to an AI step's Tools handle, one step can offer several actions for the AI to choose from.
• Endpoint fields. For job systems and accounts packages, Set endpoint fields here means you fill in the fields. Let AI populate endpoint fields lets the AI fill them, with AI instruction for this endpoint (optional) to steer it.
• Checks before writing. Writes happen as soon as the step runs. Put an Approval step in front of anything you want a person to see first, or turn on Require approval for a tool the AI calls.
• More than one account. Outlook, Gmail, Microsoft Excel, OneDrive and SharePoint can have several accounts connected to one agent. With two or more, an Account setting appears, with Primary account (default) and the others, so one step can read the support inbox while another sends from a different address.
• Per-user mode. In the agent's settings you can switch Outlook and Gmail to per-user mode, so Send email and Create draft use the mailbox of whoever runs the agent. Every other integration always uses the agent's own connection.
• Results. Each step passes on what the system returned, such as jobId, reference, contactName and status from BigChange. Find them under Step data references in the steps that follow.
• Cost: 1 credit each time the step runs, whether it reads one record or a page of 200.

The Integration group holds a step for each system an agent can work in. Open the group in the Nodes panel to see them all.

Integration19

Outlook

Sends, reads, searches and files email in a connected Outlook mailbox.

Gmail

Sends email, reads messages and manages a connected Gmail inbox.

Microsoft Excel

Reads, appends or updates rows in an Excel workbook stored in OneDrive.

Microsoft Word

Creates and updates Word documents in OneDrive.

Google Sheets

Reads, appends or updates rows in a Google Sheet.

Google Docs

Creates and updates a Google Doc with headings, lists and tables.

OneDrive

Creates folders, uploads files and downloads files in OneDrive.

SharePoint

Creates folders and uploads files to a SharePoint site library.

BigChange REST

Looks up or updates jobs, quotes, invoices and contacts in BigChange.

BigChange Web Services

Attachments, branded PDFs, signatures and other BigChange work.

Simpro

Reads and updates jobs, quotes, invoices and more in Simpro.

JobLogic

Reads and updates jobs, quotes, customers and more in JobLogic.

Sage Intacct

Reads and drafts invoices, bills, customers and vendors in Sage Intacct.

Xero

Lists and raises invoices, bills, quotes and contacts in Xero.

Ostara

Reads a work order, and can accept it, set an ETA, add notes and run job actions in Ostara.

Freshdesk

Reads and updates tickets, replies, notes and contacts in Freshdesk.

Fixflo

Reads and writes Fixflo issues, properties and people.

Cin7 Core

Reads products, stock, customers, sales and purchases in Cin7 Core.

GoHighLevel

Looks up and updates contacts, opportunities and calendars in GoHighLevel.

• Outlook: send, reply to, find and file email in a connected mailbox. For example, Send email. See Microsoft Outlook in Agent Builder.
• Gmail: the same for a Gmail inbox. For example, Reply to email. See Gmail in Agent Builder.
• Microsoft Excel: read, add and update rows in a workbook in OneDrive. For example, Append rows. See Microsoft Excel in Agent Builder.
• Microsoft Word: create and update Word documents in OneDrive. For example, Find or create document. See Microsoft Word in Agent Builder.
• Google Sheets: read, add and update rows in a Google Sheet. For example, Append row. See Google Sheets in Agent Builder.
• Google Docs: write headings, lists and tables into a Google Doc. For example, Insert blocks. See Google Docs in Agent Builder.
• OneDrive: create folders, upload and download files. For example, Upload file. See OneDrive in Agent Builder.
• SharePoint: create folders and upload files to a site library. For example, Upload file. See SharePoint in Agent Builder.
• BigChange REST: look up and update jobs, quotes, invoices and contacts. For example, List jobs. See BigChange REST in Agent Builder.
• BigChange Web Services: attachments, branded PDFs, signatures and other BigChange work. For example, Download job card PDF. See BigChange Web Services in Agent Builder.
• Simpro: read and update jobs, quotes, invoices and more. For example, Create job note. See Simpro in Agent Builder.
• JobLogic: read and update jobs, quotes, customers and more. For example, Update job status. See Joblogic in Agent Builder.
• Sage Intacct: read and draft invoices, bills, customers and vendors. For example, Create invoice, which Centrus creates as a draft. See Sage Intacct in Agent Builder.
• Xero: list and raise invoices, bills, quotes and contacts. For example, Create invoice, created as a draft unless you set otherwise. See Xero in Agent Builder.
• Ostara: read a work order, accept it, set an ETA and progress the job for your client. For example, Accept job. See Ostara in Agent Builder.
• Freshdesk: read and update tickets, replies, notes and contacts. For example, Add note, which is private unless you say otherwise. See Freshdesk in Agent Builder.
• Fixflo: read and update Fixflo issues, properties and people. For example, Get issue. See Fixflo in Agent Builder.
• Cin7 Core: read products, stock, customers, sales and purchases. For example, Get stock. See Cin7 Core in Agent Builder.
• GoHighLevel: look up and update contacts, opportunities and calendars. For example, Create opportunity. See GoHighLevel in Agent Builder.

The remaining steps have guides of their own.

• Control steps decide what happens next: Condition, Route, Resolve Entities, Dedupe, Batch, Backoff, Loop, Approval, Stop and On run error. See Conditions, loops and approvals. Each costs 0.1 credits each time it runs, apart from Loop, which is free.
• Send SMS sends an internal operational text to a UK mobile from your Centrus number. See Send text messages.
• Outbound call rings a number and hands the conversation to the Voice AI above it, and Voice AI is also the voice that answers an Incoming voice call. See Make outbound calls and answer incoming calls.
• Spike Test makes a small live read from a connected system so you can see its real field names before you build on them. See Test an agent before it goes live.

Agents spend automation credits from your organisation's monthly allowance. Every run costs 1 credit to start, and each step is charged each time it runs. A step inside a Loop is charged on every pass, so a step that runs for 40 invoices is charged 40 times.

• AI and knowledge steps (Intelligence, AI Extract Collection, File Search, Web Search) and File Parser are measured by the work done, with a minimum of 2 credits each time.
• Integration steps cost 1 credit each time they run, however many records they return.
• Extract Images, Create PDF and Geocode Postcode cost 1 credit each time they run.
• Extract Collection, Transform, Aggregate, Validate, Filter, Ingest, Merge datasets and control steps cost 0.1 credits. Loop is free.
• Tools an AI step calls are charged at the tool's own rate each time they are called, so File Search as a tool costs at least 2 credits a search.
• Browser Use is charged per browser session and also uses browser minutes. Send SMS costs 6 credits per message segment, and phone calls are charged by the minute.

In the run above, starting the run cost 1 credit, BigChange and Outlook 1 each, Create PDF 1, the Filter 0.1, the AI step 3.5 and the style guide search 2, which is 9.6 in all. Full rates, the per-run budget and the monthly allowance are in Automation credits and usage limits.

Validate in the toolbar catches most set-up mistakes in these steps before they cost you a run. An error, such as a Create PDF step with no template, stops you publishing until you fix it. A warning does not, but it usually means a step will not do what you expect.

Monday mornings

Weekly at 07:00 · Mon

Fetch last week's jobs

List jobs

Completed jobs only

No filter rules

Fetch the PO number

No goal configured

Write the summary

GPT-5.6 Terra

Build the report

No HTML template configured

2 errors

Validation

File search is enabled but no tags are selected.

Click to focus this node on the canvas

Create PDF has no HTML template configured.

Click to focus this node on the canvas

Browser Use has no goal configured.

Click to focus this node on the canvas

Filter has no rules configured, so every item will be kept.

Click to focus this node on the canvas

• Limits. Intelligence can use its tools up to 30 times per run of the step (10 by default). Browser Use sessions run for up to 30 minutes and 100 steps. Extract Images returns up to 100 pictures. A dataset from Ingest holds up to 50,000 rows for 7 days.
• Name your steps. Rename each step to say what it does in this agent, such as Fetch last week's jobs. The names appear in the Runs panel and in Step data references, and renaming never breaks a reference.
• Errors. Every step has Advanced: error & retry at the foot of Details, where On error can stop the run, carry on, or send the item down an error exit.
• Try it on real data. Use a Spike Test step, or Run now in the Runs panel, before you publish. See Test an agent before it goes live.