The Assistant — CogniWork

The assistant is the front door to a workspace. It reads and writes your data, builds views, defines workers, wires up automations, runs code against your files, and navigates you around the app. Everything it can do, you can also do by hand — it is faster, not privileged.

How a turn works

Ask for something and the assistant works in a visible loop: it looks at what already exists, decides what to change, makes the change, checks the result, and tells you what it did. You see each step as it happens, including the tools it reaches for.

It only picks up the abilities a task needs. Asking about a chart doesn’t load the file tools; asking to import a spreadsheet doesn’t load the charting knowledge. This keeps it accurate on long jobs — the cost of a bloated context is bad tool choices, and you feel that as “it forgot what I asked three messages ago.”

Steering it

Interrupt freely. If it’s heading somewhere wrong, stop it. You don’t have to wait for a turn to finish.

Correct in place. “No, group by owner, not by stage” is enough. You don’t need to re-describe the whole view.

State constraints as rules. “Never delete a record — always mark it inactive” is remembered and applied to future work, not just this turn.

Ask it to show you. “Show me the pipeline board” navigates you there. “Show me what this would look like” renders a throwaway view you can look at and discard.

Ask what it did. “What did you just change?” gets a plain summary. Useful before you walk away.

What it remembers

Memory is per workspace and survives between sessions. It holds three kinds of thing:

Facts — what your data means. That “MRR” is measured in dollars, that stage won is terminal, that Dana handles the northern region.
Rules — constraints you’ve set. Things it must not do, formats it must follow, people it must ask before acting.
Decisions — choices already made, so they aren’t relitigated. Why leads and contacts live on one type instead of two.

Memory has a sense of time. A fact recorded in March that you correct in July doesn’t fight with the new one; the older statement is superseded rather than duplicated. When something you told it is no longer true, say so — “we don’t use the webinar source anymore” — and it stops applying.

You can ask what it remembers at any time: “what rules are you working under?”

Files in the conversation

Attach a file to a message and the assistant can read it, transform it, and write the result back to your drive. Spreadsheets, CSVs, documents, images, audio. See Drive & computer for what happens to the file afterwards.

Voice

You can talk to the assistant instead of typing, and have it talk back. It handles interruption the way a person does — start speaking and it stops. Voice uses the same memory, the same data and the same tools as the text chat; it is a different microphone, not a different assistant.

Delegating

Some jobs are too big for one conversation — a full data migration, an audit of every record, a long report. For those, the assistant hands off to a worker: a focused sub-agent that goes away, does the job, and reports back while you carry on. You see its progress inline and can stop it mid-flight.

You can also do this explicitly: “give this to a worker.”

Read AI workers for the full picture.

When to take over

Do it yourself when:

You know exactly what you want. Editing one field name by hand is faster than describing it.
The change is destructive. Deleting a type or a workspace is a page you should look at with your own eyes.
You’re debugging. Opening a view’s data source and reading the query tells you more than asking why a number is wrong.

Data structures — what the assistant is actually creating.
Saved views — how the pages it builds are put together.