Knowledge and files

Use a knowledge collection for material an agent needs to search. Use workspace files for documents and other artifacts created or used during a run. A live lookup tool is the better source for facts that must be read directly from another system at the time of the request.

Create a collection

1. Open Knowledge and choose New collection.
2. Enter a recognizable Name and describe what the collection contains. Agents that find collections use this description.
3. Choose the available Searched by option. An embedding-backed setup needs a compatible embedding provider; availability depends on the environment.
4. Create the collection and add documents.
5. Wait for indexing, then try a representative search on the collection itself.

The console accepts text material: plain text, Markdown, and JSON. CSV is treated as plain text. Convert other formats into supported text before adding them; do not assume that uploading a PDF or office document here will extract its contents.

Indexing, Indexed, and Failed describe document processing. A collection with no indexed documents cannot yet supply useful passages. Fix failed documents and inspect actual search results before troubleshooting the agent's instructions.

Attach knowledge to an agent

In Agent → Capabilities → Knowledge, attach the collections the agent should normally use. The definition carries named knowledge slots. During publication, confirm which collection fills each slot in this environment.

The published version keeps that collection binding. A new collection elsewhere with a similar name does not silently replace it. Runs can also receive additional authorized collections at start through the developer interfaces; see Knowledge in the SDK.

Include instructions for when to search and what to do when the collection contains no answer. A returned passage is evidence to assess, not proof that every requested fact is present or current.

Discovery and collection permissions

Knowledge controls can permit finding other collections and creating collections. These are separate from the tools-and-skills discovery scope. Enable only the behavior the task needs, and check the associated grants and environment policy.

A collection has its own access management. People may have read or manage access, and the collection can be restricted to named agents. Giving an agent a collection binding does not grant every caller access to its contents. The Access check helps identify the missing part.

When a collection is locked to agents, maintain that allowed-agent list as part of the setup. When sharing with a person, distinguish reading material from managing documents and access.

Workspace files

Turn on Capabilities → Workspace files → Allow workspace files when a run should work with files. The model must support the required function tools. The run's file area lets people inspect the artifacts it produces.

Workspace files are working material. A knowledge collection is indexed reference material. Saving a file does not automatically make it searchable by every agent. Access to search results and access to original documents are separate concerns.

Workflow Agent steps can also expose a Shared folder setting. Use it when the process needs deliberate file sharing between steps; passing a result value and sharing a folder are separate decisions.

A useful pattern

Keep support policy in a collection. Read current account status through a lookup tool. Ask the agent to draft its response using both, and save a report in its workspace if a file is needed. Send the proposed response to a workflow Approval step before an external sending tool acts.

This keeps searchable guidance, current observations, working artifacts, and external actions understandable as separate parts of the same process.