Back office ops · Production

Dust builds synthetic filesystem navigation tools enabling AI agents to traverse company knowledge like a Unix filesystem

The problem

Dust's AI agents had no way to navigate company data by structure or hierarchy — they could only perform semantic search. When agents needed to locate a specific file by name or browse a folder's contents, they invented path-like query syntax as a workaround.

First attempt

Semantic search alone was insufficient when agents needed structure-based navigation — locating a specific entry in a meetings database by position rather than meaning cannot be reduced to a semantic query.

Workflow diagram · grounded in source
1
User query triggers agent
trigger
“User asks: "What was in the TeamOS section of the last team weekly's notion doc?"”
2
Locate content with find
ai_action
“find - Searches for files by name within hierarchies”
3
Browse directory with list
ai_action
“list - Shows folder contents (like ls)”
4
Read file content with cat
ai_action
“This lets agents read documents in chunks, jump to specific sections, and filter content, all without exploding their context windows”
5
Semantic search within subtree
ai_action
“Apply focused semantic search within those specific subtrees to understand the broader context”
6
Locate file in tree
ai_action
“Use locate_in_tree on the results to understand where related files live in the architecture”
Reported outcome

Dust shipped five Unix-inspired filesystem commands (list, find, cat, search, locate_in_tree), now live as 'Advanced Search' in the Agent Builder, enabling agents to navigate organizational data as fluently as a Unix expert navigates a filesystem.

Reported stack
listfindcatsearchlocate_in_tree
Source
https://blog.dust.tt/how-we-taught-ai-agents-to-navigate-company-data-like-a-filesystem/
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

Dust shipped five Unix-inspired filesystem commands (list, find, cat, search, locate_in_tree), now live as 'Advanced Search' in the Agent Builder, enabling agents to navigate organizational data as fluently as a Unix…

What tools did this team use?

list, find, cat, search, locate_in_tree.

What failed first in this deployment?

Semantic search alone was insufficient when agents needed structure-based navigation — locating a specific entry in a meetings database by position rather than meaning cannot be reduced to a semantic query.

How is this back office ops AI workflow structured?

User query triggers agent → Locate content with find → Browse directory with list → Read file content with cat → Semantic search within subtree → Locate file in tree.

Related back office ops cases