Talking to the agent

The composer

The input hints at it: Enter sends, /new starts a new session. Sessions inside a topic are separated by a “New session” divider.

Attachments: files from disk, drag and drop, pasting an image from the clipboard. An unsupported format is rejected with a clear message.

Attach context — a note, document, task, project, meeting, email or recording. For an email you choose what to attach: the whole thread or a single message.

Tools — a selector limiting the set for this request. Two groups: external resources (Jira, Confluence, the internet, mail sync) and local data (search over notes and documents, notes, tasks, meetings, projects, finance, documents).

Run in the background — the marked request goes to Agent chats: “started in the background — progress in Agent chats”.

While the agent works

• “The agent is thinking…”;
• Progress (N steps) — an expandable list of tool calls;
• Reasoning, when the model exposes it: “Thinking — thought 3/7…”, then “Thought (7 thoughts)”;
• Work plan — a checklist with progress;
• “Working in the background…” with a Cancel button.

On failures the agent retries by itself and shows it honestly: “retrying in N s (2/3)”.

The agent can ask right in the thread: options as buttons plus a free-text field. Afterwards the thread records “answer noted” or “question closed without an answer”.

Action confirmation

An action with consequences is shown as a card before it runs:

• what exactly will happen: creating, updating, deleting or acting on a note, task, project, meeting or subscription;
• a preview of the edit — replace, insert, delete or append with exact lines (“lines 12–18”, “at the end”);
• fields: title, content, tags, status, priority, amount, currency, category, dates, participants;
• the rationale and the sources;
• a risk marker: “safe” or “needs attention”.

Buttons: Allow, Always for this kind of action (within this conversation), Reject. Deletion warns separately that it cannot be undone.

File operations outside the workspace offer “Remember for this folder” — with the caveat that it lasts until the app closes and does not apply to writes or sensitive folders (.ssh, .env, .aws).

Result verification. After executing, the agent re-reads the result. If it diverges from what was expected you get “mismatch with reality”; if it couldn’t be re-read — “not verified”. The app doesn’t pretend everything went through.

Answer self-check

Before an answer is shown it is checked, and if the model claimed something it didn’t do, the answer is rewritten with a “Self-check” marker and a reason:

• claimed creating a note or task without a tool call;
• reported changing data without a call;
• mentioned research or analysis without real calls;
• stated an intention without acting;
• looped on repeating one phrase.

The original draft opens with Show draft — it isn’t hidden from you.

Message actions

Copy, create a task, create a note, retry the request.

When the context overflows

The Context overflow dialog offers two ways out: Delete old messages (keep the last N) or Start a new chat.

Errors worth recognising

• “the cloud provider is unavailable” — check the internet or switch the model;
• “your plan’s concurrent AI request limit has been reached”;
• “the answer didn’t form” — the agent already retried;
• a dedicated message about TLS inspection: an antivirus corrupts large requests, add the domain to its exclusions;
• “background processing is disabled” — with the reason.