What are custom Agents?
Today, Benchling AI Chat runs synchronously, so users have to directly kick off a chat to get their work done. The latest Preview feature out of Benchling AI - Agents - make it possible for users to configure a custom Agent and define when it should run, either in response to an event or on a set schedule.
The same underlying Benchling AI capabilities are supported with Agents, they just introduce a way to define a reusable and custom agent configuration.
Note: While in Preview, chats triggered by custom Agents do not consume credits. Any follow-up chats will consume credits as normal.
Get started with example Agents
- To help you get started, example Agents can be found at the top of the Agent configuration page.
- Click one of the examples to pre-populate the agent configuration, make any adjustments as needed.
- Click Save.
How to configure an Agent
Agents can be configured via Chat or manually via a form.
Configure an agent via Chat
- Start a new chat
- Describe the agent you want setup
- Confirm the proposed plan
- Optionally, review your new agent or share it with collaborators
Configura an agent manually
Go to AI Chat Settings > Agents (a modal will open)
Click Add
Specify Name, Description, Instructions.
Think of Instructions as a normal chat prompt. It will be sent to the agent when triggered. This can be edited at a later point if needed.
Configure the Trigger type (time based or event based).
For event-based triggers, we currently support a set of events for Study, Entry, Analysis, DNA Sequence and RNA Sequence.
Note: Depending on the data type selected, the supported events will populate below (eg for Entry, an agent can only be triggered on entry creation - we are working on building out support for more events. If there are particular events you’d like to see, please let us know!)
For time-based triggers, days of the week and time of day can be specified. Along with how often you want this agent to run.
Note: With this configuration, the Agent would run every Monday morning at 8am PDT.
Specify the Agent mode
In Default, agents will only be able to do research and make drafts private to you.
In Full actions, the agent will be able to edit or create Benchling objects and share its outputs with teammates.
Optionally, in Full actions mode a destination folder can be selected to limit where the agent can edit or create objects. Before making edits outside of this folder, the agent will prompt for your approval.
Click Save. The Agent will now fire automatically based on the trigger, using the same permissions as the user who configured it.
When needed, an Agent can be Paused or Archived. An Agent does not get triggered when it is Paused or Archived. Pause an agent if you would like the option to easily resume it later. Archive an agent if you no longer have a use for it.
Note: An Agent can be unpaused at any time to resume activity.
Active and Paused agents can be viewed in the Agent listing view. In the image below, the first agent is paused and the second is active.
Event-based vs. schedule-based triggers
| Event based | Time based | |
| Fires when | A specific event happens (e.g., study phase changes, an entry is created) | A fixed cadence (e.g., daily, weekly, every Monday at 9am) |
| Best for | Reactive workflows - respond the moment something changes | Proactive/routine workflows - recurring checks, summaries, reports |
| Example | "When a study phase changes from design to execution, summarize it and flag anomalies" | "Every Monday morning, pull all approved entries, review their contents against our company entry guidelines skill, and flag any that don’t follow those standards." |
Agent Permissions and Access
Permissions
Agents run as their creator (just like Benchling AI Chat). Agents do not yet have their own identity or permissions and they inherit the permissions of the user who configured them.
Access
Any user on the tenant can create an agent. After an agent is triggered, its creator can share the resulting chat with other users. To further automate the agent's work, the creator can also share the agent and its outputs automatically with teammates. See Agent Sharing for more details.
Visibility
Agent chats show up in the user's chat history. The way to distinguish a chat triggered by a custom agent vs Benchling AI Chat is via the icon tied to the chat. A robot is used to indicate when a chat was triggered by an agent and hovering over the icon indicates this.
Notifications
The best way to get notified when an Agent runs is to enable agent notification emails. We are working on supporting in-app notifications for when agents complete for a more seamless experience.
Agent Sharing
You can add collaborators to your agents and control what they're able to see and do.
Who can share an agent?
Agents currently run with their creator's permissions, so only the creator can share an agent with other users on the tenant.
Available Access Policies
Creators can grant one view-only access level:
- Viewer agent and its runs: user can see the agent configuration and automated runs of the agent (Tip: useful for sharing agent chats with teammates automatically)
Note: Creators can't currently grant edit access or the ability to manage sharing for an agent.
How do shared agents appear for added collaborators?
- Shared agents appear under the collaborator's "Shared with me" tab in AI chat settings.
- If shared with Agent viewer with runs, the agent's runs also appear automatically in the collaborator's chat listing in the AI side panel.
Coming soon
Agents will be able to run under permissions other than their creator's, and sharing options will expand accordingly.
Usage tracking and limits
While in Preview, custom Agents do not consume credits. They are powered by Benchling AI Chat, so their usage can be tracked in the Tenant admin console. To view custom Agent usage specifically, toggle on the "Include usage that doesn't deduct from credit limits" checkbox. Any dashed Chat usage can be attributed to custom Agent usage. This tracking is meant to help predict future credit consumption of custom Agents.
Usage limits
Today, Benchling AI supports 10 agents triggered every rolling 10 minutes, across the whole tenant. If more than this number of agents are triggered, any that surpass the 10 threshold will fail. Users will receive an email notifying them of any agent runs that were rate limited. We are working on an improved notifications experience.