Add knowledge, skills, and tools
Improve an agent by putting reference material, repeatable methods, and capabilities in the right places.
Log in or sign up to make these docs interactive
Choose your team, project, or agent to turn guidance into links to the exact place in your workspace.
Add context and capability after you can name the gap you are solving. A clear separation between knowledge, memory, skills, and tools keeps the agent easier to maintain.
Quick steps
- Review a failed or weak run and decide whether the missing piece is a fact, maintained source, repeatable method, or executable capability.
- Open the selected agent's Knowledge, choose Add Source, select the source type and agents, then configure and run its indexing mode.
- Add a concise memory only for a durable fact that should remain available across runs.
- Open the agent's Tools and enable the narrow skill, tool, or bundle required by the work.
- Re-run the same kind of card and verify that the agent finds the source or invokes the capability without unrelated access.
- Remove duplicate, stale, or unused context after the successful test.
1. Diagnose the gap
Review a representative card and work stream. Did the agent lack a fact, fail to follow a repeatable method, or lack the ability to perform an action?
- Use knowledge for maintained reference material.
- Use memory for a concise fact that remains useful across runs.
- Use a skill for a reusable procedure and quality checks.
- Use a tool for an actual capability or external operation.
Do not add the same material to every layer “just in case.”
2. Add a knowledge source
Open the selected agent's Knowledge page and choose Add Source. Select a website, Google Drive folder, or another available source, give it a descriptive name, and choose the agents that should use it. A Google Drive source first asks you to Connect Google Drive and select a folder.
Open Source Information and choose its indexing mode: no index for direct access only, Cache pages locally for a retrievable cache, or Full index for search and extraction. Set a page limit, choose Save Indexing Settings, then start crawling. Lowering the mode deletes data created by the higher mode, so review the confirmation before continuing.
Allow indexing to finish, then inspect the crawl status and learned preview. Test with a question whose answer is present and one whose answer is absent. The agent should find the first and remain honest about the second.
Tell the agent when the source is authoritative and how to handle conflicts.
3. Add or enable a skill
Open the agent's Skills page. Search the available skills, open the one you need, and review or edit its method before enabling it. Set it to activate always, when relevant, or remain disabled. For a new procedure, ask the agent to build a skill or create one with a clear trigger, sequence, safeguards, quality checks, and expected output; review the proposal before it becomes part of the agent's durable method.
Review any required tool bundles. Enabling a method should not silently expand the agent's authority; grant capabilities separately and deliberately.
4. Grant a tool or bundle
Open the selected agent's Tools page, search tool bundles, and enable the smallest bundle that closes the gap. Self-building and agent-administration capabilities are separate choices and should be enabled only for roles that need them. For an external operation, confirm the destination, required permission, network policy, and credential boundary. A custom HTTP tool should expose a specific approved operation, not a general-purpose tunnel.
Choose Add Tool for a custom operation, define and save the tool, then add any required destination-bound values through Save Secrets rather than in the tool description.
An agent may propose a tool it needs. Treat the proposal as a reviewable design request, not as evidence that access should be granted.
5. Re-run the same kind of work
Use a comparable card and inspect both outcome and work stream. Confirm the agent used the source or method when relevant, did not invoke it unnecessarily, and produced better evidence without unacceptable cost or access expansion.
What good looks like
Each addition has a named purpose and owner. Teammates can update a source, method, or capability without rewriting unrelated agent instructions, and the agent remains honest when the configured material is insufficient.