What the Agent Can Actually Do

The core groups

Task and project management: create, update, move, assign, bulk-edit, manage sprints, dependencies, statuses and swimlanes. This is the group everyone finds.

Clients and deals: the CRM and the pipeline — reading accounts, updating deals, managing contacts, proposing and running follow-up sequences.

Chat and discussions: workspace chat, discussion threads, notifications, message templates.

Email and calendar: connected mailboxes, calendars, scheduled and recurring email.

Reports and analytics: generating reports, portfolio data, time tracking, capacity, audit search.

Knowledge and files: the knowledge base, project files, templates.

Planning: sprint planning, capacity, horizons, estimation.

Developer and GitHub: repos, pull requests, the dev surfaces.

Workflows and automations: automation rules, recurring tasks, background jobs.

Finance: accounting integration — invoices, transactions, ledger.

Administration: governance, SLA policies, support, the undo stack.

The connection rule

Here is the thing that explains most 'why can't it do that' questions.

Tools for a provider you have not connected are not offered to the agent at all. They are not present-but-failing; they are absent. If Gmail is not connected, the agent has no Gmail tools and will tell you it cannot read your mail — which is correct, and which changes the moment you connect it.

The same applies to opt-in features. Community app tools, the remote browser, exec meetings and several others are off by default, and while they are off the corresponding tools are stripped rather than merely refused. A workspace that has not opted in sees an agent that behaves as though the feature does not exist, which is deliberate: nobody should be shown a capability nobody asked for.

So when the agent says it cannot do something, the first question is not 'how do I phrase this better' — it is 'is the thing connected or enabled?'.

Beyond the workspace

The agent has web search and can read web pages. These are genuinely non-redundant capabilities: search finds things, page reading gets the actual content of a specific URL.

That makes 'research this company before the call and put a summary in the deal notes' a single request rather than three tools and a copy-paste.

It can also generate images and other media, and it can produce charts and diagrams inline when a picture is the clearer answer. Asking for a chart of something in your workspace is a reasonable request rather than a stretch.

What it genuinely cannot do

Being precise about limits is more useful than being encouraging about capabilities.

It cannot see systems you have not connected. There is no inference from absence: a project living in a tool you have not linked simply is not visible.

It cannot exceed the governance you have set. At safe autonomy the risky tools are removed from its toolset entirely, so it is not choosing to decline — it has nothing to decline with.

It cannot cross the workspace boundary. Another workspace's data is not visible even to you-as-the-same-account.

It cannot unsend. Anything that has left the building is gone, which is exactly why external sends default to requiring your approval.

And it cannot make your decisions. It can surface the trade-off, draft the options and tell you what the charter says. Which one you pick is not a capability question.

Finding out what yours has

The fastest way to learn your own agent's reach is to ask it. 'What can you do with our Google Workspace connection?' produces an accurate answer for your workspace, which a written list cannot, because the answer depends on what you have connected and enabled.

One caveat worth knowing: the agent occasionally under-reports its own capabilities — having a tool is not the same as knowing you have it. If it declines something you are fairly confident it can do, naming the capability directly usually resolves it.