Files and Voice Just Became First-Class Citizens in Your AI Harness
Atlas now lets you drag, drop, and discuss—images, docs, spreadsheets, and voice—directly in conversations with NyLi, Rhonda, and Atlas Owner AI.

Key takeaways
- Files and voice are now first-class participants in Atlas conversations—no more copy-pasting or context switching.
- Multimodal context is governed: small files inlined, large files chunked with summaries, and Rhonda’s access strictly gated.
- Voice is built for real work: Deepgram default with Gemini fallback, transcription and text-to-speech, and a design that doesn’t foreclose real-time bidirectional voice later.
- This isn’t about adding features. It’s about making the harness smarter, so the intelligence can actually work with your business context.
You shouldn’t have to flatten your work to fit the AI
Most AI tools treat files like attachments in an email—something you reference, not something the system actually understands. You upload a PDF, the model reads the first few pages, and you’re left hoping it didn’t miss the footnote on page 47. That’s not how work happens. You don’t send a spreadsheet to a colleague and say, ‘Here, figure it out.’ You walk them through it. You point at the cells. You argue about the assumptions. You update it together.
Atlas now lets you do exactly that. Drag a file into a conversation with NyLi, Rhonda, or Atlas Owner AI, and the model gets the full context—not just the metadata, not just a preview, but the actual contents. Images, PDFs, DOCX, spreadsheets, CSVs, even audio. The harness handles the rest: sniffing file types, scanning for threats, deciding whether to inline the full content or chunk it with a summary, and retaining access across later turns. No more ‘Sorry, I can’t see that file anymore’ after three messages.
Voice works the same way. Hit record, speak your question or instruction, and the system transcribes it (Deepgram default, Gemini fallback) and treats it like any other input. No voice on Rhonda—because anonymous surfaces shouldn’t have that kind of access—but NyLi and Atlas Owner AI can now hear you. The design leaves room for real-time bidirectional voice later, but today, it’s about making the conversation richer, not just faster.
✦ What ‘first-class’ actually means
1. **No context loss**: The model retains access to the file across the entire conversation, not just the first turn. 2. **Hybrid context**: Small files are inlined in full; large files are chunked with a summary always inlined. 3. **Governed access**: Rhonda gets images and small docs only, with hard caps and short retention. 4. **Magic-byte sniffing**: Every file is checked for type and threats before the model sees it. 5. **Conversation-scoped**: Attachments stay tied to the conversation; they don’t auto-promote into org-wide Knowledge.
The harness doesn’t just pass files—it decides how to handle them
Here’s the thing about multimodal context: it’s not just about letting the model see more. It’s about letting the harness decide how to handle that ‘more’ in a way that’s useful, secure, and efficient. You don’t want a 500-page contract inlined in full—that’s a token bonfire. You don’t want a sensitive spreadsheet floating around in an anonymous chat. And you definitely don’t want the AI to forget what it just read three messages ago.
Atlas makes those decisions for you. Small files? Inlined. Large files? Chunked, embedded, and summarized. Rhonda? Limited to images and small docs, with Turnstile-gated access and short retention. Every file gets magic-byte sniffing to confirm its type, and real antivirus scanning happens on the anonymous Rhonda surface. The result is a system that can actually work with your business context, not just pretend to.
Voice follows the same principle. The transcription and text-to-speech layer is abstracted behind one provider, so you’re not locked into a single vendor. Deepgram is the default, Gemini is the fallback, and the design explicitly leaves room for real-time bidirectional voice later. But today, it’s about making the conversation richer—letting you speak naturally, not just type faster.
- Small files (<10MB): Inlined in full for the model to reference across the conversation.
- Large files: Chunked and embedded, with a summary always inlined for context.
- Rhonda: Limited to images and small docs, with hard caps and short retention.
- Voice: Transcription (speech-to-text) and text-to-speech, with Deepgram default and Gemini fallback.
This isn’t about adding features—it’s about removing friction
Most AI tools add multimodal support like they’re checking a box: ‘Look, we can read PDFs now!’ But if the system doesn’t actually understand the file—if it can’t retain context, if it can’t govern access, if it can’t decide how to handle large vs. small files—then you’re just adding more work for the user. You’re making them flatten their work to fit the tool.
Atlas does the opposite. It removes the friction of working with files and voice by making them first-class participants in the conversation. The harness handles the heavy lifting: deciding how to process the file, retaining access, applying permissions, and even cleaning up after itself. The result is a system that can actually work with your business context, not just around it.
That’s the difference between a model and a harness. A model gives you intelligence. A harness makes that intelligence operational. Files and voice aren’t just new inputs—they’re new ways for the harness to connect that intelligence to your work.
✦ The economics of richer context
Token usage is a cost metric, not an ROI metric. The real question is whether richer context leads to better outcomes: fewer errors, faster decisions, higher-quality work. Atlas’s AI Value Indicators track exactly that—work completed, time returned, cycle-time reduction, recommendation acceptance—so you can measure the value of multimodal context, not just the cost.
See it on your own data.
Connect your tools and Atlas shows you what matters.
Frequently asked questions
Can I use voice in real-time conversations, like a phone call?
Not yet. Today’s release supports transcription (speech-to-text) and text-to-speech, so you can record a voice message and have the AI respond in kind. The design explicitly leaves room for real-time bidirectional voice later, but that’s a separate phase.
What happens to files after the conversation ends?
Attachments are conversation-scoped, meaning they stay tied to that specific conversation and don’t auto-promote into org-wide Knowledge. Rhonda’s access is further limited by hard caps and short retention. The goal is to keep context available where it’s needed, without creating a sprawling, ungoverned knowledge base.
How does Atlas decide whether to inline a file or chunk it?
The harness uses a hybrid approach: small files are inlined in full for the model to reference across the conversation, while large files are chunked and embedded, with a summary always inlined for context. The exact thresholds are configurable, but the default is designed to balance token efficiency with usability.
Keep reading
Related resources
Why Your AI Agents Keep Hitting the Same Wall—And How to Fix It
BlogWhen Your AI Strategy Is Just a Model with Nowhere to Go
BlogPast the Spreadsheet: Making AI Accountable, Not Just Another File
BlogMeasuring Work Differently in the Age of AI
BlogAtlas Now Speaks Files, Images, and Voice—Without Losing the Plot
BlogQuestioning the Work Itself: What Should Even Exist to Automate
Newsletter
The consolidation memo.
Practical insights on AI, operations, and the future of business software. No fluff.