Doe does not require you to choose a working mode. You ask for the outcome, and Doe decides how much agent work is needed.
For simple requests, Doe can respond directly. For larger requests, Doe can plan the work, gather context, coordinate subagents when useful, validate the result, and produce finished artifacts for review.
You can keep the request lightweight or delegate a complete task. The interaction stays the same: describe what you want done, add the context Doe should use, and review the result.
How Doe Routes Work
| Request type | What Doe may do |
|---|
| Quick question or rewrite | Answer directly and keep the exchange lightweight |
| One file summary | Read the file, cite relevant sections, and answer in the conversation |
| Multi-file research with citations | Search, read, compare, cite, and synthesize a structured result |
| Spreadsheet creation or cleanup | Create, edit, format, validate, chart, export, or sync a workbook |
| Cross-app work across email, CRM, chat, and files | Gather signals across connected apps and produce a reviewable result |
| Data analysis, charts, or generated files | Use analysis tools, create artifacts, and save outputs for review |
| A recurring task | Help turn the prompt into a Loop that runs on a cadence |
For bigger outcomes, be explicit about sources, output format, approval rules, and validation. Doe will decide the right depth of work.
What Happens During Agent Work
Doe keeps the experience simple: you describe the outcome, then review the work. Behind that, a larger request moves through visible stages.
| Stage | What you see |
|---|
| Plan | An execution plan card with checkable steps; larger plans pause for your confirmation before work starts |
| Gather context | Activity lines as Doe searches files, apps, prior sessions, websites, or uploaded material |
| Work | Named subagents and tool actions as Doe reads, analyzes, edits, drafts, and creates |
| Review | A verification pass that can request revisions before the answer is finalized |
| Synthesize | The result, with sources attached inline |
The Activity Stream
Substantial work narrates itself in the conversation. Activity is grouped into collapsible sections, and each finished section settles into a one-line receipt with what happened and how long it took, such as Searched web for pricing · 12s. Expand any section to see the individual actions, their targets, and their results.
Subagents appear as named rows with their assignment. When a subagent finishes, it files a completion receipt you can open: scope coverage (such as 4/4 covered or blocked), a confidence level, and structured Findings, Issues, Gaps, and Recommendations, so you can see not just what was found but what was not.
Sources You Can Open
Answers cite their sources inline. Each citation is a small pill; hover to preview the source and why it was cited, and click to go to the exact evidence:
| Source | Clicking the citation |
|---|
| Web page | Opens the source in a new tab |
| File | Opens the file preview at the cited page or location |
| Spreadsheet | Opens the workbook at the exact cited cells |
| Word document | Opens the document at the cited text |
| Calculation | Shows the code and the calculation step that produced the number |
| Reasoning | Shows the reasoning step behind the claim |
Verified Before It’s Done
Doe checks its own work before presenting it:
- Deliverables must exist. If the result claims a saved or downloadable file, the run is checked for real persistence evidence before it can complete. A clickable file pill only appears once the file has actually been saved.
- A reviewer pass runs on substantial work. You will sometimes see
Reviewing response or Requesting revisions in the activity stream as Doe rejects and reworks its own output.
- Loops verify their hard requirements against the Loop’s spec on every run. See Loops.
Built for Long Runs
- Transient provider issues retry silently. Momentary overloads and rate limits are absorbed with automatic retries and fallbacks; you only hear about a problem if it persists.
- Paused work keeps its progress. When a run pauses, for an approval or at your request, everything done so far is preserved, and resuming picks up where it left off.
- Interrupted runs keep their trail. Stopping a run keeps the full activity record, sources, and any artifacts produced so far.
- Large data is never silently truncated. Oversized results are externalized into complete, navigable artifacts, and heavy processing moves into the code sandbox, so a 50,000-row export comes back whole.
Memory of Prior Work
Doe can look back at previous sessions and audit its own past runs. Ask things like:
Doe searches prior conversations, folders, and date ranges, and can trace any past claim back to the exact action and source that produced it before answering.
From One Request to Repeatable Work
Start with one delegated outcome. If the same work should happen again, turn it into a repeatable task.
| Pattern | Example |
|---|
| One-off request | Build a market map of 50 competitors and create a cited workbook. |
| Parallel subagents | Research each competitor's pricing, positioning, integrations, and customers in parallel. |
| Reviewable action | Draft follow-up emails for these accounts. Do not send until I approve. |
| Scheduled Loop | Every Monday at 9am PT, update the pipeline risk report and draft a Slack summary. |
Subagents
Subagents are specialized AI assistants that help with focused parts of a larger task. They are useful when a request has independent pieces of work, such as researching several competitors, reading many files, comparing multiple apps, or validating a spreadsheet.
You do not need to configure subagents directly. Doe uses them when they help finish the work faster or more reliably, and the activity view shows the work at a high level.
Examples:
Doe can divide the research by competitor or topic, then combine the findings into one table.
Doe can search the documents, read the relevant sections, and create a structured spreadsheet for review.
Doe can compare signals across connected apps and produce a prioritized action list.
Managing Multiple Agent Runs
You can start more than one substantial request and move between conversations while work continues. Active work is visible in the sidebar and preview pane. Open any run to inspect progress, review files, or send a steering message.
Useful follow-ups while agent work runs:
Focus only on enterprise accounts.Use the file I just uploaded as the template.Skip LinkedIn and use HubSpot plus Gmail instead.Pause before sending anything externally.
What Agents Can Use
Depending on your permissions and connected context, agents can use:
- Uploaded files, Library files, folders, and prior sessions
- File-content search across indexed documents
- Document extraction and PDF/DOCX editing
- Spreadsheet creation, formatting, formulas, charts, validation, and export
- Dashboards and published static sites as durable deliverables
- Library organization: creating files and folders, renaming, and moving work into place
- Saved memories and approved lessons from previous work
- Connected apps such as Gmail, Slack, HubSpot, Salesforce, GitHub, Linear, Notion, and more
- Web research, website review, and structured data extraction
- Database analysis through connected database integrations
- Code sandboxes for calculations, analysis, charts, and generated artifacts
- Loops for scheduled recurring work
Task Playbooks
See practical ways to deploy agents
Chat and Preview Pane
Watch activity, files, spreadsheets, and sources beside the conversation
Research and Code Sandboxes
Learn how agents research, analyze data, and create artifacts
