Create and use AI agents

An agent is a saved assistant for recurring work. It combines instructions, a model, tools and, optionally, its own knowledge files. Visible features depend on your role, workspace and instance configuration.

Agent library

Start an agent#

1. Open Agents.
2. Search or filter the library. Recently used provides quick access.
3. Open View details to check the description, example prompts, tools and model.
4. Select Use. Form agents open structured input instead.

My Agents contains agents you created or added. Select Add agents to discover agents released by your organisation and add them to your personal list; this does not change their shared definition. Templates contains approved starting points. Copying a template creates a separate agent of your own.

Create an agent#

Select New agent. The builder has seven numbered steps: Identity, Instructions, Knowledge, Tools, Model, Team and Overview. Editing an agent you own initially opens Overview — Step 7 of 7. The header switches between Build and History; Expert mode and YAML open technical fields and the configuration view.

Start with a narrow task. State the task, inputs, process, boundaries and output format. For example:

Compare two versions of a contract.
Use only the files provided.
List changed duties, deadlines and liability terms in a table.
Flag ambiguity rather than making assumptions.

Optimise can rewrite a draft. Review the suggestion before accepting it. The technical view also provides YAML preview, download and import.

Agent builder overview and seven configuration steps

Example instance with sample data. Available controls depend on your role and configuration.

Configure the seven steps#

1. Under Identity, maintain name, description, category, tags, icon, gradient, conversation starters and input type. A form suits runs that always need the same information; set a label, help text, type and required status for each field. User fields hold personal values per person.
2. Under Instructions, define role, process, boundaries and output format. Treat Optimise and templates as starting points and review the text.
3. Under Knowledge, attach files, choose retrieval mode and, where needed, maintain the context document.
4. Under Tools, grant built-in capabilities, integration actions and sub-agents.
5. Under Model, choose the fixed model, reasoning effort and offered run parameters.
6. Under Team, review visibility and members. Configure delegation and sub-agent procedures under Tools.
7. Overview shows the configuration and open issues before finishing, saving or publishing. Its change links take you back to the relevant step; a completed step marker is not a quality assessment.

Models, tools and knowledge#

The model list contains only models offered by your administrator. Reasoning controls appear only where supported. Greater reasoning effort can take longer and consume more balance.

Grant only required tools. Common entries include File search, Web search, Deep research, Code interpreter and Generate documents. Other capabilities include Translate document, Base64 decoder, Update context files and Ask the user. Availability depends on feature switches, services and approvals. Integration tools also need an enabled integration and, where applicable, your personal connection. Under Confirmation, an unticked write action asks before every run; an individually allowed action can run without asking.

Tool picker

Attach files under Knowledge and wait for Ready before relying on document search. A changed chunking strategy may require re-embedding. The AGENT.md context document suits short durable guidance and is read in full on each run.

User fields hold personal details such as a signature or region for each person. Required fields must be completed before the first run. Personal account bindings are not copied.

Sharing, versions and templates#

• Private: you and explicitly added members.
• Project: access follows the assigned project.
• Admin (global): visible to everyone; administrators alone can set it.
• Internal (admins only): hidden from ordinary users.

Available choices depend on your role and workspace mode. Editing can create a draft; colleagues keep using the published version until you publish the draft. History lets you compare and, where offered, restore versions.

Preview makes real model calls and consumes balance. Test confirmation-gated actions in a normal chat afterwards.

Publish as template submits an agent for administrator review. Verified and Featured are editorial gallery labels, not certification.

Form input and personal user fields#

Under Identity, select Form for Input type, then Add field. Types are Text, Long text, Select, Number, Email, Checkbox, Date and File. Give each field a name, optional description and, for Select, its options. Mark mandatory input Required and use the move actions to order fields. A form run checks required fields before sending; files use the ordinary upload path.

Form fields describe one run's input. User fields instead keep personal defaults across all chats. They support short and long text, Markdown, HTML and documents. The Field ID is the stable prompt key and cannot change after saving. You can set a default value, maximum length and, for documents, a maximum count. HTML is shown only through a sanitised preview.

Under Model, choose a model that supports the task and files you intend to use. A model retained by an existing agent may appear even when it is no longer offered for new choices. An unavailable model needs an administrator or agent editor to choose a replacement.

Expert mode also exposes response length, creativity/temperature and Top P controls, along with the tool-step Recursion limit. The latter bounds the agent's repeated reasoning and tool steps; it is not the number of people or specialists. Reaching a run or output limit can leave the task incomplete. Start with the instance defaults and change one setting at a time while testing. Model/provider support determines which parameters take effect.

Under Team, inspect existing access. Adding a Member grants use of a finished agent immediately; membership of an incomplete agent only takes effect after it is finished. An Admin invitation additionally permits editing files and the system prompt, but must be accepted and expires after the displayed period. Accept the invitation before using the agent through the normal sharing controls. Promotion to Admin is not offered afterwards: remove and invite the person again in that role. Personal integrations always come from the person running the agent.

Deleting a shared agent removes access for other members, and the interface warns before doing so. Copy previews which tools, files, sub-agents, forms, contexts and templates travel, and which personal bindings do not.

File handling and document templates#

Under Knowledge, Expert mode exposes File handling. This setting controls the route for uploads:

This setting does not prevent a model from using its general knowledge. If answers must rely on approved sources, say so in the instructions and check the resulting citations. A bypassed file will not become searchable merely because it was attached.

Chunking settings control how documents are divided for search. Keep inherited settings unless you have a concrete retrieval problem. After changing them, use the offered re-embedding action for existing knowledge and wait for processing to finish.

Agent document templates belong to Generate documents and use templates available on the instance. The context document AGENT.md is different from indexed knowledge: it supplies instructions or compact reference material in full. Keep it focused, and check whether you are editing the live context or a draft copy.

Delegate to sub-agents#

Open Tools and add a specialist through the action picker or the sub-agent section. The main agent must be able to use the selected specialist; granting a name does not bypass its access rules.

• Give the specialist a distinct job and explain when the main agent should call it.
• Use the return contract to specify the expected result, such as findings with sources and unresolved questions.
• Set dependencies only when one specialist genuinely needs another's output. A pipeline expresses a fixed sequence.
• Review per-specialist overrides and the overall delegation limits. The main agent and its specialists consume balance during a run.
• Keep write authority explicit. A specialist's access to tools does not by itself authorise every external change.

Test a small task first, then expand the sub-agent cards in chat to see what ran and what each specialist returned. If a specialist is already referenced by another agent, deletion can be refused until that reference is removed.

Evaluate an agent#

Where Evaluations is offered, create or import cases with an input, expected output and, where useful, expected tools. Run the dataset and inspect individual cases rather than relying on an aggregate score. Evaluation runs use saved agent configuration; they do not automatically test your open unpublished draft. Use the builder preview and an explicitly opted-in chat to test draft changes. Evaluations make real model calls. A passing run proves only the examples tested.

Save, test and publish changes#

A new agent and an unpublished draft are different states:

1. Enter a name in Identity and select Next. This creates an incomplete agent; simply opening the builder does not create one.
2. While the agent is incomplete, changes save automatically. Check the Saving…, Saved or Not saved indicator. Knowledge, tools, team and overview need that initial saved agent before their controls are available.
3. Work through the steps and resolve blocking checks under Overview. Select Finish to make the agent usable. An unfinished agent is not ready for members to run.
4. Once finished, edits to a private agent require Save Agent. Saving a shared agent prompts you to use a draft or Save live anyway, because an immediate save affects its users.

Select Create draft for a finished agent when you want to prepare changes separately. In a draft the action becomes Save draft; Use this draft in my chats opts only you into testing it. Publish replaces the live version for everyone and switches testers back to the published version.

History compares published versions with another version, the current builder configuration or the open draft. Restore to draft allows review before publishing. Restore and publish now creates a new version at the top; it does not rewind history. Publish or discard an open draft before restoring another version. Config-driven agents are read-only and are managed through deployment configuration.

Troubleshooting#

• Check filters, My agents and project membership if an agent is missing.
• Open Integrations when a connected tool requests sign-in.
• Check Ready and the File search grant when knowledge is missing.
• Publish the draft when colleagues should receive a change.
• Ask your administrator if a model, service or tool is not offered.

Next: Files · Projects · Integrations · MCP tools