An agent needs both a clear job and the means to do it. A skill can teach a method; a tool performs an operation; a connection supplies access to an external service. These are separate parts of setup.
Match the missing piece to the problem
| Part | What it supplies | Example |
|---|---|---|
| Agent instructions | Role, working rules, expected decisions, and boundaries | Research product claims and flag unsupported statements |
| Skill | Reusable written guidance for a type of work | A repeatable content-review checklist |
| Tool access | Operations the agent is allowed to invoke | Read project documents or use the browser |
| App connection | An authorized account or service connection | A connected Google account or MCP service |
| Knowledge source | Material available for retrieval in the relevant scope | An approved product reference |
| Execution context | The current project, space, and, where relevant, codebase or computer | The repository attached to this project |
An installed skill does not connect an app, grant tool permissions, or make a missing file available. Likewise, connecting an app does not necessarily enable its tools for every agent.
Ask the agent to diagnose access first
Use a concrete request:
Before starting this task, check which tools and connections you can actually use. Tell me which required capability is missing, whether a connection or tool permission is needed, and what I can do next. Do not make account or permission changes yet.
The current runtime includes capability inspection and tool discovery. An agent can search and load tool groups it is already authorized to use. Loading makes those tools visible to the agent; it does not expand permission. Seeing Search Tool Groups or Load Tool Group in activity is therefore not evidence that a new connection was created.
Connect through the conversation
Ask the agent to show a connection card for the app the task needs. It can also prepare a custom MCP connection card from verified public server settings. Select Connect, review the setup, authorize the intended account or server, then check the returned connection status.
The card is the user-controlled setup handoff, not a tool grant. Current source requires the requester’s eligible organization role and an agent with connection-request capability. A missing card can therefore be a role or capability issue rather than an unsupported service.
For custom MCP, a successful connection exposes a connection-specific tool option. The agent still needs that option granted. For an existing native app, missing provider permissions or a connection that needs reconnection can lead back to a connection card, while a healthy connection with missing agent access needs a separate tool-capability grant.
Use this diagnostic request:
Check whether this is a missing or unhealthy connection, missing provider scope, missing agent tool access, or a source unavailable in this space. Show the appropriate connection card if setup is needed and allowed. Do not change permissions or perform external writes yet.
Keep passwords, tokens, and other credentials in the dedicated setup form. Do not send them in chat. An authorized person can review any required agent capability request or use the agent’s Tools settings.
Add a reusable method
For a procedure that should apply again, ask for a named skill rather than repeating a large prompt every time:
Create a reusable skill called Evidence-first content review. It should require a source for each product claim, preserve uncertainties, check audience and tone, and return a draft for review rather than publishing. Attach it to this agent and tell me whether it was created or an existing skill was reused.
The source-backed skill tool creates organization-owned Markdown guidance and attaches it to the current agent. An existing name can be reused rather than overwritten. Ask explicitly before replacing shared guidance; other agents may rely on it.
To revise a method, name the skill and the change. Ask the agent to confirm the saved skill and attachment, not merely quote a proposed replacement in chat. Current source says changes apply on subsequent runs. Skill creation and editing have not been exercised in the documentation demo.
Choose the configuration surface carefully
Authorized agent settings expose instructions, skills, tools, and knowledge. The full agent workspace and compact chat settings dialog have different layouts. Some accounts can use an agent but cannot open a particular administration page; see Roles and access.
The space assistant has role-defined capabilities. It is not simply a specialist with every tool enabled. Use Space assistant to understand what it coordinates, and Project manager to decide where execution belongs.
Troubleshoot in the right order
- Confirm the intended agent and project.
- Ask the agent to inspect its available capabilities.
- Check whether the required account is connected.
- Check the agent's granted tool access.
- For source-backed services, check whether the relevant source is attached to the current space.
- For coding or computer work, check the selected codebase or execution environment.
- Retry a bounded read before attempting a consequential write.
Do not assume every integration uses the same resource model. Some use an organization connection; others require selected sources. The provider-specific guides remain the authority for Gmail, Yahoo, Drive, Calendar, and LinkedIn boundaries.