Build your first agent

A step-by-step guide for people with no technical background. You will set up an agent from a template, try it safely with a dry run, turn it on and check its work. Then you will build one from scratch and learn how to write instructions that work. For the full reference, see AI agents.

What an agent is, in one minute

An agent is a job you describe once. SourceLace then does it for you, on a schedule or when something happens.

• It reads your systems as you, with your own access. It can never see more than you can.
• It can only do the actions you list: draft emails, make a document, propose record changes or post in Slack or Teams.
• Emails are always drafts in your own mailbox. SourceLace never sends email.
• Record changes wait for your approval, unless an admin allowed otherwise.
• Every agent starts as a draft. A draft can only do dry runs, which change nothing.

You find agents under Agents in the menu. If you see "You can't create agents", ask an admin: your plan or your organization's settings decide who can.

Start from a template

A template is a ready-made agent for a common job. It is the fastest and safest way to start.

1. Go to Agents and press Start from a template (or open the Templates tab).
2. Search or pick a category. Each card says what the template does, who it is for, when it runs and what it needs.
3. Look at the pill at the bottom of the card. Your systems fit means you can use it now. Needs a system you don't have yet means an admin must add a source first.
4. Press Use template. A short wizard with four steps opens.

Worked example: Stale opportunity nudges

Say you are a sales manager at Corvanta. Deals in Salesforce go quiet and nobody notices. You want a short note to each rep every weekday, listing their quiet deals and asking for a next step.

Step 1. Systems. The wizard asks which of your systems to use for each need:

• CRM: choose Corvanta Salesforce.
• Mail (for drafts): choose your Gmail or Outlook. This is where the drafts appear.

If a system shows (connect it first), sign in to it on Data Sources, then come back.

Step 2. Questions. The template asks a few things. Keep the defaults for your first try:

Step 3. When and how. The schedule is Weekdays at 09:00. The time zone is filled in from your browser; change it if needed. Under Actions you see "Draft a nudge to an opportunity owner" with Drafts only, nothing is sent.

Step 4. Name and results. Give it a name, such as "Corvanta stale deals". Under Where results go, keep A new project named after the agent. Each run's summary becomes a chat in that project. Press Create agent.

You land on the agent's page with the message "Created as a draft. Try a dry run, then turn it on." Now go to Try it with a dry run.

Build one from scratch

Use Agents → New agent when no template fits. The editor has five parts.

The job. Give a Name, a short Description and the Instructions (see Write instructions that work). Skill to follow (optional) lets the agent follow one of your organization's skills, such as a house style for QBR decks.

Systems it may read. Tick only what the job needs. The agent cannot read anything else. Tick your mailbox, calendar, or Slack or Teams here too if an action or trigger uses them; otherwise saving fails with "add it to the agent's sources first".

What it may do. Add actions with the buttons + Change records, + Draft emails, + Make a document (in the results) and + Post in Slack or Teams. See Choose systems and actions.

When it runs. Pick Only when I press Run, On a schedule, Before calendar events or When something happens. See Decide when it runs.

Where results go and Advanced. Pick a Results project so summaries become chats you can ask follow-up questions in. Keep the default limits at first; they are under Advanced, with When a run fails (you can also get an email draft in your own mailbox).

Press Create draft. Every later save is Save as a new version, and each run records which version it used.

Write instructions that work

Write instructions as you would brief a new colleague on their first day. They are smart, but they do not know your company, your fields or what "good" looks like.

Checklist for good instructions

• What to read: which system and which records ("open opportunities in Salesforce").
• Which ones count: the filter, with numbers ("no activity for 21 days or more", "amount 10,000 or more").
• What to compare or work out: totals, changes since last week, what is missing.
• What to produce: drafts, a document or a summary, and what each one contains.
• How much: a limit ("at most 15 drafts", "the 10 largest").
• What NOT to do: the things a new colleague might get wrong.
• How to finish: a short summary with the key numbers.
• Short numbered steps. One step, one thing.

Example 1: vague goal

Before:

Keep an eye on the pipeline and tell people to update their deals.

After:

1. In Salesforce, find open opportunities with an amount of 10,000 or more and no activity for 21 days or more.
2. Group them by owner. For the 15 owners with the largest total, draft one email to the owner listing their deals (name, stage, amount, close date, days since last activity) and asking for the next step and its date.
3. Do not put one owner's deals in another owner's draft. Do not guess next steps.
4. Finish with a summary: how many quiet deals, their total amount and how many drafts you made.

Why it is better: it says what to read, which records count, what to produce, how many and what not to do.

Example 2: no comparison

Before:

Tell me about support tickets.

After:

1. In ServiceNow, count open incidents by priority and by assignment group.
2. Compare incidents opened in the last 7 days with the 7 days before, by priority, and say what went up or down.
3. List priority 1 and 2 incidents older than their target (4 hours and 1 day), with number, short description, group and age.
4. Make a one-page document with these tables. Do not include incidents from the "Test" group.
5. Finish with three lines: total open, the biggest change and the oldest priority 1.

Why it is better: "tell me about" becomes counts, a comparison, a list with fields and a clear output. Note that each run starts fresh and does not remember earlier runs, so compare within the data, not with "last time".

Example 3: too much power

Before:

Fix the account data in our CRM.

After:

1. In Salesforce, find accounts of type Customer with an empty Industry or an empty Billing Country.
2. For each, look only at what Salesforce already holds, such as the mailing country of the account's contacts. Only when the answer is clear, propose a change to Industry or Billing Country. Never change any other field.
3. At most 10 accounts per run. If you are not sure, list the account in the summary instead of proposing a change.
4. Finish with a list of accounts changed, accounts skipped and why.

Pair this with a Change records action on Account that may only Update records for the fields Industry, BillingCountry, with I approve each change.

Why it is better: "fix" is open-ended. The new version names the fields, sets a limit, says when to hold back and keeps every change waiting for your approval.

Tips

• Use the names of things as your system shows them: stage names, object names, group names.
• Say who drafts go to and what the subject should be.
• If text in a record or email tells the agent to do something, it is ignored. The agent treats all data as data. You do not need to write rules about that.
• The instructions cannot give the agent more power. If an action or system is not listed, it is refused whatever the instructions say.

Choose systems and actions

Least access. Give the agent only the systems and actions the job needs. A pipeline report needs to read your CRM and make a document. It does not need to change records or draft emails.

The four actions:

Drafts or record changes? Start with drafts, documents and summaries. They change nothing in your systems, so they are safe to try. Add record changes only when the drafts look right for a few weeks. On the Team plan, agents make drafts and documents only.

Approvals. Keep I approve each change for anything that changes records. "Without approval" only works if an admin has allowed that object in Agents → Settings, on a plan that includes it. If not, saving tells you so.

Decide when it runs

Only when I press Run. Good while you are still testing.

On a schedule. Choose How often (Every hour, Every day, Weekdays, Every week or Every month), the time At (HH:MM) and the Time zone. For weekly, pick the day. For monthly, pick a day from 1 to 28 or Last day of the month. 08:00 stays 08:00 through daylight-saving changes.

Before calendar events. Choose the Calendar (your Gmail or Outlook), the words in Events whose title contains (such as QBR) and Hours before the event (48 is two days). The match ignores capital letters. The agent runs once for each matching event, and cancelled events are skipped. The run is told which meeting it is for.

When something happens. The agent checks a system every few minutes and runs once for each new record it finds, such as each new priority 1 case. Think of it as a saved search that SourceLace repeats for you:

• You write the search in the system's own query language, with {since} where the "since when" time goes. For example: SELECT Id, Subject FROM Case WHERE Priority = 'High' AND CreatedDate > {since}.
• Check every (minutes): 5 at the most often.
• Most runs per check: if many records appear at once, the rest wait for the next check.
• Each record starts one run, once. To run again each time a record changes, fill in Run again when this field changes (optional), such as LastModifiedDate.
• When you turn the agent on, or save a new version, records that already exist are noted and skipped.

If you do not know the query language, start from the Sev 1 escalation or Closed-won handoff template. They already have the right search for each system.

Try it with a dry run

On the agent's page, press Dry run. It starts within a minute. A dry run reads your systems and plans everything like a real run, but it changes nothing and drafts nothing.

Open the run from the Runs list on the agent's page. The run page shows:

• Result: Done, Failed or Stopped at a limit, marked dry run. A note says "Dry run: nothing was changed or drafted. Below is what it would have done."
• Summary: what the agent found and what it would have done. Check the numbers against what you know.
• Changes: one card per draft, message or record change it would have made, marked dry run. Each shows the fields with their value Now and After. For the stale deals example, check the To, Subject and Body of each draft.
• Steps: each tool it used, Where (system and object), how many Rows came back, the Outcome and the Time. Refused means it tried something it is not allowed to do. Failed gives the reason, such as a wrong field name.
• Credits: what the run used. Multiply by the runs per day to see the monthly cost.

What to look for:

• Did it read the right records? Zero rows often means a filter or name is wrong.
• Are the numbers right? Compare with a report you trust.
• Are drafts going to the right people with the right content?
• Any Refused or Failed steps? Fix the instructions or add the missing system.

Edit and dry-run again until you are happy. It usually takes two or three tries.

Turn it on

Press Turn on. The status becomes On and Next run shows when it will run. Your plan sets how many agents can be on at once; if all are in use, pause one first.

Read runs and results

• The Agents page lists your agents with their status, when they run, the last run, the share of runs that succeeded in the last 30 days, credits this month and changes waiting for approval.
• Click an agent to see its setup, its instructions and all its runs.
• When a run has a summary or files, it becomes a chat in the agent's project. Use Open the results chat on the run page to ask follow-up questions.
• The run page never shows the data the agent read. That data stays in your systems.

The approvals inbox

Agents → Approvals lists every change waiting for you. The tab shows how many are waiting.

• Each card says what would change (such as "Update Opportunity …"), which agent and run proposed it, why, and until when it waits.
• The table shows each field's value Now and After.
• Press Approve and apply or Reject.
• Approving checks the record again first. If it changed since the proposal, the change is marked stale and nothing is written; the next run proposes again.
• Changes not decided within 7 days expire.
• All shows past decisions too.

Pause or stop an agent

• Pause skips the schedule and triggers until you press Turn on. You can still run it by hand.
• Stop is for when something is wrong: after asking you to confirm, nothing runs, not even a dry run, until its owner turns it on again.
• Delete agent (under More on the agent's page) removes the agent with its runs and waiting approvals.
• Admins can stop every agent at once in Agents → Settings. Runs in progress stop at their next step.

Common mistakes and fixes

Checklist before you turn it on

• [ ] The name and description say what it does, so others understand it.
• [ ] The instructions say what to read, which records count, what to produce, how many and what not to do.
• [ ] It reads only the systems the job needs.
• [ ] Its actions are the fewest that do the job. Drafts and documents before record changes.
• [ ] Record changes list only the fields needed, and wait for approval.
• [ ] Slack or Teams posts list only the right channels.
• [ ] The schedule, time zone or calendar words are right.
• [ ] For event agents, the search finds the right records and uses {since}.
• [ ] A dry run finished, read the right records and the numbers look right.
• [ ] The drafts or changes in the dry run are what you would have done yourself.
• [ ] The credits per run, times runs per day, fit your budget.
• [ ] The results go to a project you will check.
• [ ] You know where to look: Approvals, Needs your attention and the results chat.

Once the agent works well, you can share it: see Make templates for your team.