FAQ - Agent templates Private Preview
Common questions about agent templates in Atlan, covering access, security, cost, run behaviour, and what's currently in the works.
Common questions about creating and running agents from a template. For what an agent does and what it writes, see Understand agent templates.
Access and permissions
Who can create and run agents?
Workspace admins and governance admins. A run-only role, so that stewards can run an agent an admin configured, is in the works.
How do you get access?
Agent templates are enabled per tenant during private preview. If you don't see a Create agent button in Agents studio, raise a support request.
Who can see what agent produced?
What an agent publishes is visible to anyone who can see the glossary it writes into, governed by Atlan's existing persona and purpose controls.
Run detail, including the action log and the citation snippets an agent read from your sources, is reached through Agents studio. Because Agents studio is currently limited to admins and governance admins, only those roles can open it.
Sources and security
Can agents write to source systems?
No. Both integrations are read-only and request no write scopes. An agent reads from your source and writes only into Atlan.
What can agents read?
Exactly what the connection's credential can read, and nothing more. That credential is your access-control lever, which is why a dedicated service account scoped to the spaces or repositories you want in play is recommended.
Can users see source content they don't otherwise have access to?
Agents read through one admin-scoped, read-only credential rather than as the person who triggered the run, so agent output isn't access-intersected per user. What an agent writes into Atlan is governed by Atlan's existing persona and purpose controls.
Which sources can you connect today?
GitHub and Confluence. See Knowledge sources for setup, and in the works for what's coming.
Can two Confluence sites or two GitHub organizations be connected?
Not currently. Atlan holds one connection per source, per workspace, and connecting a second credential for a source replaces the one in use for every agent.
Output and review
How does agent avoid creating duplicate terms?
Through its skills, using Atlan as a source in its own right. Before creating anything, the agent queries your workspace through Atlan's MCP server to discover what the glossary already holds, then decides per candidate whether to create a new term or enrich an existing one.
The same discovery step is what stops an agent contradicting your team: if a term already carries a definition somebody wrote, the agent doesn't replace it. Where a candidate is close to something that exists, the run reports it rather than creating a near-duplicate alongside it.
Does agent overwrite what team has written?
Agents write additively. They're instructed and gated not to replace a definition, term link, or README authored by a person: they fill what's empty and refresh only content they generated themselves.
Every value an agent writes carries the isAIGenerated tag, so you can find and audit agent-written content anywhere in Atlan and confirm what has changed.
Are terms published as drafts for approval?
No. Terms are published live and tagged isAIGenerated, so you can find and review everything an agent has written. Review happens in Atlan's normal glossary flow. Triggering a governance workflow on agent-written changes is in the works.
Can agents undo what they published?
Not automatically. There is no un-publish, no rollback, and no bulk delete of what an agent created, and deleting the agent doesn't delete its output. Run a sample first: it publishes nothing and shows you what a full run writes.
Where to review what full run wrote?
On the run itself, which lists the terms with their citations and offers a CSV export, and in the glossary through View in Atlan.
Does sample run write anything?
No. A sample run is constrained by the agent's run contract, which prohibits every write path, and by the publishing step, which doesn't execute in sample mode. Atlan also validates each run's report and fails any sample run showing evidence of a write. Tool-level isolation of sample runs is in the works.
Why does everything in sample show as new?
A sample skips the duplicate check against your existing glossary so that it can return quickly. A full run performs the check properly and may enrich existing terms rather than creating new ones. Treat a sample as a signal about your instructions, not a prediction of create versus enrich.
Runs and scheduling
How long does run take?
A sample run takes a minute or two. A full run depends on how much it has to read. A scheduled run that doesn't finish within its allotted time is stopped, and anything it published before stopping stays in Atlan.
Can two agents run at once?
Different agents can run independently. A single agent runs one at a time, and both run buttons are unavailable while that agent has a run in flight.
Can I stop run?
Yes. Stop appears on the run's stats card while a run is in flight.
Stop is a termination rather than an undo. The run ends at its next step, so it may continue for a moment after you confirm, and anything it had already published stays in Atlan. If a full run wrote terms you didn't want, delete them yourself; there is no un-publish and no automatic cleanup.
A stopped sample run leaves nothing behind, because a sample publishes nothing.
What happens if scheduled run fails?
It waits for the next scheduled slot. Scheduled runs aren't retried automatically, so a transient source failure means one missed run rather than a retry storm.
Why did scheduled run not happen?
If a run is still going when the next one is due, the new one is skipped rather than queued. Fewer runs than your cadence implies usually means the previous run was still working.
Why did run finish quickly with no output?
Usually the agent might reach the source but couldn't read from it. Connecting a source verifies that the server is reachable; it doesn't yet verify that the credential has been granted read access to the tools an agent needs. If a run completes quickly with little or no output, check that the credential can actually read the repositories or spaces in the agent's scope, then run again.
The same symptom appears if the source was disconnected after the agent was created.
Run found nothing worth extracting. Why does it show as failed?
An empty result is currently reported as a failure rather than as a successful empty run. Check the action log to see what the agent read before assuming something is broken.
Configuration and cost
Can skills be added to agents?
Yes. On the Skills step you can search the skills your workspace has already written and tick the ones this agent must carry, or create a new one by uploading a skill folder or a .zip. Platform skills ship with the template, are always bound, and aren't listed.
A skill is a reusable, named capability that tells an agent how to perform a task. Because skills are assets in Atlan, a skill you write once can serve many agents. Use one for something the template can't know, such as a naming convention or a review procedure. Changing which skills an agent carries is recorded as a new version.
Can I edit agent after creating it?
You can change run instructions, scope, sources, and skills. Name, description, and the template are fixed at creation, because the agent's name is also the name of the glossary it writes into.
Why do some changes create new versions?
Changing sources or skills changes what the agent can reach, so it's recorded as a version. Changing scope or run instructions retargets it within what it can already reach and takes effect on the next run without a new version.
Does every agent need unique name?
Yes, and it matters more than tidiness. An agent's name determines the glossary it writes into, so two agents sharing a name write into the same glossary. Give every agent a distinct, specific name at creation, because the name can't be changed afterwards.
What does run cost?
Nothing during private preview.
Afterwards, runs are priced per outcome: you are charged for each item an agent creates in Atlan, not for the run itself and not for how long it takes. The rate depends on the kind of outcome, and different templates produce different outcomes.
Run detail lists exactly what a run produced, so you can size a full run from a sample before committing to it.
Which model does agent use?
Runs execute on Atlan's managed AI infrastructure. Model choice isn't configurable.
Template needed isn't listed. Can it be requested?
Yes, and there are two ways, which do different things.
- The bell on a Coming soon template card registers your interest in that specific template. Those signals decide which pipeline gets built next.
- Looking for another agent? Tell us what you need, at the foot of the template picker, is for a use case that's not in the catalogue at all. Describe the outcome in your own words: what the agent reads, and what it writes into Atlan. That reaches the Atlan team directly.
Concrete requests are more useful than abstract ones. "Read the SharePoint SOPs and tell me which of the near-identical copies is the golden one" tells Atlan more than "support SharePoint."
In works
Agent templates are in private preview and the following are actively being built. Feedback on priority is welcome through your Atlan contact.
| Area | What is coming |
|---|---|
| More knowledge sources | SharePoint, Glean, and Google Drive, plus direct file upload for teams who prefer not to open a live connection |
| Custom MCP servers | Connect an MCP server your organization runs, alongside the sources Atlan ships |
| More templates | Knowledge Curator, Code & API-Spec Lineage Agent, and Skill / AOP Compiler. Select the bell on a template card to register interest and influence the order |
| A run-only role | So that an admin can configure an agent and a steward can run it |
| Governance workflow triggers | So that agent-written changes can start your existing review workflow |
| Version restore | Rolling an agent back to an earlier version from the Versions tab |
| Tool-level isolation for sample runs | Enforcing the no-write guarantee at the tool boundary rather than through the run contract |
See also
- Understand agent templates: what each template writes and where it lands
- Create an agent from a template: the six-step setup flow
- Run and schedule an agent: runs, settings, and cadence
- Knowledge sources: connect GitHub and Confluence
- Product release stages: what private preview means
Need help
If you need assistance after reviewing this page, contact Atlan support.