Read Agents overview first. It covers the configuration surface this page assumes. Agents aren’t available on free plans.
The five parts
An agent that behaves incorrectly is usually misconfigured in one of these five parts. Output at the wrong time indicates a trigger problem. Output in the wrong shape indicates an instructions problem.
Give each agent one job. An agent that tags data, writes a summary, and notifies an owner will do all three less reliably than three separate agents would. If the instructions contain “and then also”, the second half belongs in another agent.
Digital Twins are also agents, but they run continuously as a listener you talk to in Chat rather than as a job that runs and finishes. The trigger guidance below doesn’t apply to them.
Choose a trigger
Work that happens on a cadence uses a schedule. Work that responds to a change in the workspace, such as a tag being applied or new data arriving, uses a Dovetail event trigger. The agent runs immediately, with the changed item as its input. The trigger determines how quickly a problem is raised. An event trigger reports an issue on the day the data arrives. A weekly schedule reports it at the next run. A schedule set to run more often than the data changes produces empty output.Write instructions and attach skills
Instructions have the largest effect on output quality, which determines whether a result can be used as it is or rewritten first. Describe the outcome rather than the procedure, because step-by-step instructions fail when the input changes shape. Writing instructions describes what a strong instruction contains. Attach skills for anything the agent needs to look up, such as a tag taxonomy, a report template, or an internal glossary. Each skill is injected in full at run time, so several short docs perform better than one long one. Facts that apply company-wide belong in workspace context docs, which every agent picks up.Scope tools and connectors
Every Dovetail tool is enabled by default. Disable the ones the agent doesn’t need, to limit what it can change and because fewer options produce more accurate execution. Connectors are disabled until you enable them. Linear, Notion, Hex, Salesforce, Slack, Canva, Gmail, Lemlist, and custom MCP connectors are available. With a connector enabled, an agent can act on what it finds. An agent that identifies a tagged security concern can create an urgent issue in the Linear queue the engineering team already reviews, so the item enters the existing triage process without anyone transferring it manually.Agents can deliver output to Slack, Microsoft Teams, and email. Invoking an agent from Slack or Teams isn’t available yet. Mention the agent with
@ in Chat instead.Test before going live
An agent that writes to your workspace or messages colleagues acts under your account, so test it before pointing it at a shared destination.1
Run it on demand two or three times
Read the output each time and adjust the instructions.
2
Send output to yourself first
Use your own doc or email address before a shared destination. Formatting is the most common problem.
3
Test a period with little activity
Confirm the agent reports that there is nothing to report, rather than producing unsupported content.
4
Open the sources it cites
Check that the citations lead to the underlying data.
Review output after launch
Read the output for the first few scheduled runs and assess whether it is still useful, not only whether the agent ran. An agent producing the same generic summary each week is no longer worth its schedule. When quality declines, change one element of the instructions and run it again rather than rebuilding the agent. Only the creator can edit an agent, and ownership can’t be transferred. The Agents FAQs cover this and how to pause a schedule.Results
A configured agent produces recurring reports without anyone writing them, and raises issues while they are still current rather than at the next scheduled review. Because each output cites the data it drew on, anyone questioning a finding can open the interview or ticket the agent used and assess it directly.Where to go next
Use case library
Example prompts by team, with the connectors each relies on.
Writing instructions
Patterns for digests, taggers, and summarizers.
Dovetail tools
What an agent can do in your workspace.
Digital Twins
Agents built from what a segment has said, that you talk to in Chat.