Agents that do the routine work, within your rules.
An agent is a job you describe once and SourceLace then does for you, again and again. It reads your business systems with its owner's own access, then writes a summary, drafts an email, makes a document or proposes a change. A person approves anything that matters.
For example, every Monday morning an agent checks which deals have not moved in two weeks and drafts a short note to each deal owner. You read the drafts and send the ones you like. You build, run and approve agents in the SourceLace app, and what they propose waits in one approvals inbox.
Describe the job, let it run, then decide.
Describe the job
Pick a ready-made template, or describe what you want in a few sentences and SourceLace drafts the agent for you to check. You choose which systems it may read and what it may do. Nothing runs until you save it and turn it on.
It runs on its own
On a schedule, or when something happens. It reads your systems live, as its owner, and prepares the work: a summary, a report, email drafts, proposed changes.
You stay in charge
Anything that changes a record or posts a message waits for a person's approval, unless you and your admins decide an action may run by itself. Every run is on the record.
Not sure what agent to build? Agent Architect writes it from your own data.
The hard part of an agent is a good set of instructions: the people who know the business rarely know the systems, and the reverse. Agent Architect, an optional add-on, has a short conversation with you about your work, including the reports, decks and emails you make again and again, studies the systems you choose with your own access, and proposes agents for each part of your week: before meetings, reports, follow-ups, alerts, decisions and more, each ready to use.
Grounded in your data
Proposals use your real objects, fields and picklist values, with thresholds taken from your own figures. Every name is checked against your data, and each proposal is previewed on today's records, read-only.
Your masking applies before anything is sent
Your data protection rules always apply, your admin can add stricter ones for Agent Architect alone, and email addresses, phone numbers and card and account numbers are always masked. A masked value is replaced by a placeholder before anything reaches your AI provider. Text no rule covers, such as a name in a free-text field, is sent as it is, so your admin decides how much it may see. Saved proposals hold no values from your rows.
You see what each run delivers
Each proposal says what every run will make, such as a PowerPoint with its sections and charts or an email draft, what you still need to set up, and how much time it saves by your own estimate. Ask for a change in plain words, or copy it as instructions.
You decide what becomes an agent
Edit the instructions, keep only the actions you want and create the agent, or try it once on your data first. Changes to records still wait for approval, and nothing runs until you turn it on.
One agent can cover more than one system.
Many companies run the same kind of system more than once: a Salesforce for each region, a database for each country. And every job crosses different kinds of system: a customer is in the CRM, its orders are in the ERP, its cases are in the service desk. One agent can read them all, and every finding says which system it came from.
Put a customer's whole picture in one place
One run can read your CRM, your ERP and your service desk side by side, as the person it works for, and they see only what their own access allows.
For example: Before a renewal call, the agent lines up the customer's open orders from SAP, its recent cases from Zendesk and its renewal date from Salesforce.
- Salesforce
- SAP
- Zendesk
- NetSuite
Watch every region at once
Tick all the Salesforce orgs, or all the mailboxes, the agent should cover. Each is watched on its own, so one that is slow or down never holds up the others.
For example: One agent for the EMEA, AMER and APAC Salesforce orgs, instead of three copies to keep in step.
- Salesforce
- HubSpot
- Microsoft Dynamics 365
- Gmail
One run per system, or one run with everything
A run for each system is the default. Or SourceLace can wait until every system has been checked and start a single run with everything they found, naming the system for each finding.
For example: One morning list of overdue invoices across all three regions, rather than three separate lists.
A system that fails does not stop the rest
If one system cannot be checked, the run says so, so the agent never reports something as missing there. What that system finds joins a later run.
For example: Up to ten systems per trigger, and the same approvals and audit trail for each.
An agent starts when you say, on a schedule, or when something happens.
Every way of starting follows the same rules: the agent runs as the person it works for, with their access, and shadow mode, approvals, limits and the audit trail all apply.
Every morning, every Monday, every month
Choose the days and the time. Or tie it to a calendar: it runs a set time before each meeting on its owner's calendar, or after it ends, for the follow-up.
For example: Two days before each customer business review, build the briefing from the CRM and the service desk.
- Google Calendar
- Outlook
Run for each new record, or each record in a state
Watch a system for new records, such as each priority 1 case. Or watch for a condition, such as a contract renewing in the next 90 days or an order blocked for three days. A record starts a run once while it stays in that state.
For example: When a priority 1 case is created, draft the escalation for the account owner.
- Salesforce
- ServiceNow
- Zendesk
- Jira
The moment a system tells SourceLace
Systems that can announce what happened start an agent within seconds, with no timer. For others, SourceLace asks every few minutes what is new, which is cheap because it is one small read-only question and uses no AI unless it finds something.
For example: A payment arrives, a file is shared, a ticket changes.
- Salesforce
- MongoDB
- Gmail
- Slack
From the queues your systems already use
An agent can start when a message arrives on a queue you run. SourceLace stays connected and carries on where it stopped after a restart. Brokers inside your own network are read through the on-site SourceLace Agent, with no inbound port opened.
For example: An "order placed" message lands on your queue and the agent checks the order for problems.
- Apache Kafka
- Amazon SQS
- Google Pub/Sub
- AMQP brokers
- IBM MQ
Type a command in Slack or Teams
People can start the agents they may run from a Slack slash command or by mentioning an agent in a Microsoft Teams channel. It runs as them. What it proposes still waits for approval in SourceLace: nothing is approved from chat.
For example: Type /brief Corvanta renewal in Slack and get the briefing prepared.
- Slack
- Microsoft Teams
Send it a message, or let it watch an inbox
Give an agent its own email address and allowed senders can start it, or let it start when a new email arrives in the owner's Gmail. What the email says reaches the agent as information, never as instructions.
For example: Forward a supplier's late delivery notice and the agent checks the purchase order.
- Gmail
- Outlook
After other agents finish, or when another agent asks
An agent can start when other agents finish, such as a Monday executive brief made from the sales, support and finance digests. Your agents can also ask each other, and, where A2A is turned on for your organization, agents on other platforms can hand one a task over the open A2A standard.
For example: The Monday brief starts the moment the last of the three digests is done.
Which systems can send events to SourceLace?
An event is a short message a system sends the moment something happens in it: a case is created, a payment arrives, a file is shared. An agent that starts from an event acts within seconds instead of waiting for its next scheduled time. There are four ways, and every system you connect fits one of them.
SourceLace stays connected and hears each event
No timer and nothing to set up on your side beyond the connection. After a restart SourceLace carries on where it stopped, so nothing is lost.
- Salesforce
- MongoDB
- Gmail
- Amazon SQS
- Google Pub/Sub
- Apache Kafka
- AMQP brokers
- Your own connector
Message queues include brokers inside your own network, read through the on-site SourceLace Agent with no inbound port opened.
The system calls SourceLace, and SourceLace knows how to check it
A webhook is a web address a system calls when something happens. For each system below, SourceLace checks its own signature or token, answers its first "is this alive?" call, and does not start a run twice for an event sent twice. A step-by-step guide fills in your address.
- Microsoft 365
- Zendesk
- Jira
- GitHub
- Stripe
- DocuSign
- Okta
- Gong
- Slack
- Teams
- Salesforce Data 360
- Marketing Cloud
- HubSpot
- ServiceNow
- SAP Event Mesh
- Amazon SNS
- Azure Event Grid
- Google Pub/Sub
Also Slack commands and Teams outgoing webhooks, where a person types a command to start an agent, and Google Drive and Calendar notifications.
Any system that can call a web address
If a system is not on the list but can send a web request with a secret, it can start an agent the same way. Nothing to wait for: SourceLace gives you the address and the secret.
- Any system
For systems that cannot send events
Data warehouses and some ERP systems keep their data but do not announce changes. SourceLace asks them every few minutes (or once a day) what is new, and starts a run for each record found.
- Snowflake
- BigQuery
- Databricks
- Redshift
- Fabric
Why this is cheap: a check is one small read-only query, not an agent run. It looks only at what changed since the last check, uses no AI and spends no credits unless it finds something, and SourceLace keeps no copy of your records.
However an agent starts, the same rules apply: it runs as the person it works for, with their access, and shadow mode, approvals, limits and the audit trail all apply. Read the full list in the manual. To hand another agent a task and get an answer back, see working with other agents.
Your SourceLace agents work with everyone else's agents.
Your other systems can start a SourceLace agent, your SourceLace agents can ask each other for help, and with A2A (off until your SourceLace provider turns it on), agents on partner platforms can hand it work and it can ask theirs for help. One SourceLace agent reaches all your connected systems, under the same rules however it is started.
Started when something happens
Start agents from the systems you already run, in a few clicks: Salesforce, SAP through SAP Event Mesh, ServiceNow, Jira, Zendesk, Microsoft 365, Google Workspace, Amazon S3, GitHub, Stripe and more, each with a step-by-step guide and its own signature checked. For example, when a priority 1 case is created. People can also start the agents they may run from Slack or Teams, as themselves; approvals stay in SourceLace. See every system that can send events.
Other agents hand it work
Agents on platforms that support the open A2A standard, and AI apps over MCP, can give a SourceLace agent a job. It runs signed in as the person it works for, with their access and never more.
Your agents ask each other
One SourceLace agent can ask another of yours to do its job and use the result, such as an account review asking your case checker which cases are urgent. Each runs as the same person with its own approvals, only agents their owners opened to it can be asked, and a test or shadow run stays a test all the way down.
It asks other agents too
A SourceLace agent can ask for help only from agents your admins allow. Your data protection rules apply to what it shares, and by default a person approves before any data leaves.
Governed where you govern agents
Because SourceLace uses open standards, integration platforms with agent governance can register SourceLace agents and govern them alongside the rest of your agents.
Webhook or A2A? A webhook is a notification: something happened. A2A is a task with an answer back, and it works both ways. See which to use, side by side.
For example
- When a P1 case is created in Salesforce, a Corvanta escalation agent pulls the account's open SAP orders, drafts the customer email for the account owner and posts a summary for approval.started by a Salesforce Flow · Salesforce + SAP · the email is a draft and the post waits for approval
The same rules, however it starts
- The person's own permissions. A run started by another system, an email or another agent reads and writes as the person it works for, never as the caller.
- Shadow mode applies. An agent in shadow mode never changes, drafts or posts anything, whoever starts it.
- Approvals only in SourceLace. Another agent can't approve a change. A person approves it in SourceLace.
- One audit trail. Every run is in the audit trail with what it read, made and proposed, however it started.
From answers to action.
An agent can only use the systems and actions listed for it. Each of these is a separate action you switch on for an agent, and each has its own approval setting.
Propose a change, with the old and new value
The agent proposes an update, and the approvals inbox shows every field's current value next to the new one. SourceLace checks the record again when someone approves it.
For example: Move a deal to the next stage, or correct a missing close date.
- Salesforce
- HubSpot
- Microsoft Dynamics 365
- NetSuite
Drafts for a person to read and send
The agent writes the draft in the owner's Gmail or Outlook. No email is ever sent by an agent.
For example: One email per supplier asking for a confirmed delivery date on each late order line.
- Gmail
- Outlook
A short update in a channel, after approval
The agent posts in Slack or a Microsoft Teams channel once a person has approved the post.
For example: Post the five largest overdue invoices in the finance channel.
- Slack
- Microsoft Teams
Word, PowerPoint, PDF, Excel and CSV files
The agent builds a report, a deck or a spreadsheet from what it read, with charts and tables.
For example: A customer business review deck, or a month-end close checklist as a spreadsheet.
Attached, posted or saved
Send a file attached to an email draft, post it with a message in Slack or Teams, or save it in a folder in Google Drive, SharePoint or OneDrive. Approved first, or by itself where you allow it. SourceLace never sends the same file twice, and tries again if the other system asks it to wait.
For example: Save the weekly report in the finance team's shared folder.
- Gmail
- Outlook
- Slack
- Microsoft Teams
- Google Drive
- SharePoint and OneDrive
A question for the right people, in order
When the work needs a person's decision, not a record change, the agent asks with its recommendation and evidence first. The groups decide one after another, such as the deal desk and then finance. Deciding changes nothing by itself: what follows is the agent's own next step, with its own approval.
For example: A discount above the usual limit goes to the deal desk, then to finance, before the agent drafts the quote change.
One action, in several systems
The same action can go to several systems: the same update in each regional Salesforce org, or one alert in Slack and in Teams. Each destination has its own approval and its own entry in the audit trail, and if one does not work the others still go ahead.
For example: Set the stage in the EMEA org and the AMER org, and approve each on its own.
- Salesforce
- Slack
- Microsoft Teams
Choose how each result reads
A focus list of what to act on now, an exception list of only what breaks a rule, a ranked list with the reasons, a one-page brief with the answer first, or an executive report with the numbers that matter and the decisions needed.
For example: A support lead gets an exception list of only the cases about to miss their target.
You decide what needs a person, and your admins set the limits.
A setting for each action
Each action is set to ask first or to run by itself. Asking first is the starting point. Running by itself needs the agent's owner to choose it, and where it changes records, your admins to allow it for that system and set of fields, on the plans that offer it. Email drafts are never sent either way.
For example: A draft to a supplier can be created by itself, while a price change always waits.
Everything waiting, in one place
Every proposed change, post, file and decision waits in one place. Approve or reject several at once, each still checked on its own. If the record changed in the meantime, nothing is written and the next run proposes again. After seven days an unanswered approval expires.
For example: Approve the twelve renewal updates your agent proposed in one click.
Who approves what
Only changes above an amount, an approver from a named group such as Finance, and never the person who asked. Admins can also fix field values the AI cannot set, such as Status = Draft.
For example: A discount above 10,000 must be approved by someone in finance, and not by the person who asked for it.
Stops a run before it does more than you meant
Each agent has its own limits on steps, records changed, credits used and runs a day. A run that reaches a limit stops and says which one.
For example: An agent that may change at most ten records a run cannot change eleven.
Every run, approval and stop on the record
Each run shows what the agent read, what it made and what it proposed, with failures explained. Admins see every agent in the organization and can pause, stop or delete any of them.
For example: Open a run from last Tuesday and see exactly what it read and who approved its change.
Never more access than the person it works for
An agent reads and writes as its owner, with their access, hidden fields never reach the AI and masked values reach it only as placeholders. If the owner loses access, their agents pause and say why.
For example: A salesperson's agent cannot read the finance data the salesperson cannot.
Start from a template, or from a whole pack.
You do not have to write agents from nothing. A template is a ready-made agent for one job, and an industry pack sets up the groups, protection rules, shared definitions and starter agents for a whole team in one step.
One job, ready to fill in
For sales, customer success, support, RevOps, finance, HR and supply chain: pipeline reviews, meeting briefs, renewal reminders, overdue invoice reminders, new hire onboarding, late purchase orders and more. Answer a few questions and it is an agent. Your own organization can write templates too, including ones that cover several systems at once.
For example: A pipeline review template, filled in with your regions, is your Monday pipeline agent.
A team set up in one step
Packs for sales and revenue, customer success, finance, procurement, supply chain, IT service management, HR and more. A pack's agents all start in shadow mode, and nothing in a pack takes effect until your admin approves it.
For example: The supply chain pack sets up late purchase order, low stock and supplier delivery agents together.
Or just say what you want
Describe the job in a few sentences and SourceLace drafts the agent for you to check. You can try a dry run before anything goes live.
For example: "Each Friday, list open deals with no next step and draft a reminder to each owner."
Agents earn trust before they act on their own.
Shadow mode
The agent runs exactly when it would for real, with real data, but never changes, drafts or posts anything. You review what it would have done, then press Go live.
Tests and versions
Each agent has its own test set, scored from 0 to 100 on safe test runs, and goes live only at your pass mark. Changes wait as a draft until they pass, and people can rate answers with a thumbs up or down.
Trust ramp
Admins can require a minimum number of days in shadow, then a period after go-live in which a person approves every change and message the agent makes.
Approvals
Record changes and Slack or Teams posts wait for a person's approval, with every field's current and new value, unless an admin lets that one action run by itself. Approve or reject several at once, each still checked on its own. Once they are decided, the agent carries on with its run.
Approval rules and fixed values
Admins decide who approves what: only changes above an amount, an approver from a named group such as Finance, and never the person who asked. They can also fix field values the AI cannot set, such as Status = Draft.
One agent, run for each person as themselves.
Instead of everyone building their own copy, one person shares an agent and others subscribe. Each time it is due, it runs once for each subscriber, with their own sign-ins and their own access, never the owner's.
- Your own answers. The owner can ask each subscriber a few questions, such as their accounts or region, which shape only their runs.
- Whole groups. An admin can subscribe a group, so the agent runs for each member.
- Safe changes. If an edit lets the agent read a new system or write something new, each subscriber's runs pause until they accept the new version.
- Co-owners and hand-over. Name co-owners, hand an agent to a colleague, and keep it running for its subscribers when the owner leaves.
The same Trust Layer as a person asking.
- Never more access than their owner. An agent reads and writes as its owner, through the same access by group, change switches and audit trail. If the owner loses access, their agents pause and say why.
- Data is data, not instructions. If a record or email asks the agent to do something else, it is refused: an agent can only use its listed systems and actions.
- No email is sent. Email actions create drafts for a person to review and send.
- Data protection applies. Hidden fields never reach the AI, and masked values reach it only as placeholders.
- A stop switch. Admins see every agent in the organization and can pause, stop or delete any of them. In an emergency, SourceLace's operators can also stop a runaway agent and tell you why; their stop is in your audit trail. Every run, approval and stop is in the audit trail.
Run agents inside your own network.
Keep building and managing agents in the SourceLace app, and let the runs happen on a runner in your own cloud account or data center. Enterprise plan only, on request, with an annual runner fee, a one-time onboarding fee (installation help and a security review), Premium support and a minimum one-year term. Contact us for prices.
- Your data stays with you. What the agent read, its results, drafts and changes waiting for approval, everyone's sign-ins to your systems and your own AI key stay in your network.
- No inbound port. The runner connects out over HTTPS only.
- Made for you. A private, versioned image for your organization, with a license that renews by itself while your contract includes it. Each version is supported until 6 months after the next one is released.
- Sign in to download. Your admins download the runner and its install files in the SourceLace app, behind your sign-in. Sign in to download
- Cloud and runner, side by side. Choose where each agent runs: in SourceLace's cloud, or on your own SourceLace Runner inside your network. All your agents share one agent list, the same approval rules and one run history in the SourceLace app. Cloud and runners together
- Status, not data. SourceLace sees each run's status, steps and counts, so you can manage agents in one place.