thanks! you’ve absolutely nailed it—the team that wins this category will be the one that cracks security + UX together.
on security: your point about permission boundaries is spot-on. our current approach is deliberately conservative:
• agents only generate new documents (no editing of source files)
• source library: read-only mount
• agent workspace: isolated per user
• every file operation logged in the conversation
this solves ~80% of the risk surface for our initial use case (repetitive analysis: merging invoices, generating reports, summarizing documents). the read-heavy workflow means we can be much more permissive without the scary “accidental deletion” scenarios.
that said, we’re not claiming this is sufficient long-term. pre-flight approval, granular permissions, and audit trails are all on the roadmap.
on UX: this is where it gets really interesting. chat-as-interface has become the default (for good reason—natural, low learning curve), but i suspect we’re still in the “skeuomorphic” phase of AI interaction design.
just like early iOS apps looked like physical notepads, we’re making AI “look like slack.” but what comes after chat? chat + artifacts like Claude? what’s the interaction paradigm that’s native to agentic AI, not borrowed from messaging apps?
i don’t have the answer yet, but i think it involves:
• showing agent “work-in-progress” (not just final outputs)
• making delegation feel more natural (less prompting, more tasking)
• ambient awareness of what agents are doing (trust through transparency)
strategic tension: there’s an interesting debate in our team. i believe the wedge is automating the repetitive 80% of white-collar work—the boring, time-consuming stuff that’s currently burning 3 hours of an analyst’s day. clear roi, measurable value.
others argue that’s “not sexy enough”—that agents are better suited for creative work where they can augment human thinking, not just execute rote tasks.
honestly? i think the creative use cases are more fun to talk about, but the repetitive automation is what people will actually pay for right now. the creative stuff can be the cherry on top, not the whole sundae.
curious where you land on that: automate the boring stuff first, or go straight for the creative augmentation?
Good luck getting people to sign up to read your story!
Thanks for the advice. I am new to hacker news so… Here’s my post: https://blog.gbase.ai/blog/claude-code-for-office-workers/
Hi HN!
There's a quiet revolution happening in AI: agents are gaining the ability to actually USE computers, not just talk about using them.
We gave our AI agent filesystem access. It changed everything.
- Before: Advice "You should use pandas to merge those Excel files, then create a pivot table..."
- After: Action Agent opens files, writes Python, generates Excel, saves to disk "Here's your summary report: /cache/analysis_2025.xlsx"
Here's what we've done： - File system access (read/write from user's document library)
- Office document parsing (Excel → structured data for AI)
- Code execution (agent writes & runs Python)
- Output generation (creates Excel, charts, reports)
- Sandboxed environment (Docker isolation for security)
This isn't just about "productivity tools." It's about the *fundamental shift from AI-as-advisor to AI-as-coworker*.
The moment AI can: - Access your files - Execute code - Generate outputs
...is the moment it can start replacing repetitive white-collar work.
[dead]
thanks! you’ve absolutely nailed it—the team that wins this category will be the one that cracks security + UX together.
on security: your point about permission boundaries is spot-on. our current approach is deliberately conservative: • agents only generate new documents (no editing of source files) • source library: read-only mount • agent workspace: isolated per user • every file operation logged in the conversation
this solves ~80% of the risk surface for our initial use case (repetitive analysis: merging invoices, generating reports, summarizing documents). the read-heavy workflow means we can be much more permissive without the scary “accidental deletion” scenarios.
that said, we’re not claiming this is sufficient long-term. pre-flight approval, granular permissions, and audit trails are all on the roadmap.
on UX: this is where it gets really interesting. chat-as-interface has become the default (for good reason—natural, low learning curve), but i suspect we’re still in the “skeuomorphic” phase of AI interaction design.
just like early iOS apps looked like physical notepads, we’re making AI “look like slack.” but what comes after chat? chat + artifacts like Claude? what’s the interaction paradigm that’s native to agentic AI, not borrowed from messaging apps?
i don’t have the answer yet, but i think it involves: • showing agent “work-in-progress” (not just final outputs) • making delegation feel more natural (less prompting, more tasking) • ambient awareness of what agents are doing (trust through transparency)
strategic tension: there’s an interesting debate in our team. i believe the wedge is automating the repetitive 80% of white-collar work—the boring, time-consuming stuff that’s currently burning 3 hours of an analyst’s day. clear roi, measurable value.
others argue that’s “not sexy enough”—that agents are better suited for creative work where they can augment human thinking, not just execute rote tasks.
honestly? i think the creative use cases are more fun to talk about, but the repetitive automation is what people will actually pay for right now. the creative stuff can be the cherry on top, not the whole sundae.
curious where you land on that: automate the boring stuff first, or go straight for the creative augmentation?