Back office ops · Production

Meta builds a swarm of 50+ AI agents to map tribal knowledge across large-scale data pipelines

The problem

Meta's AI agents failed to make useful edits on a large proprietary pipeline because they lacked any map of the codebase's tribal knowledge — non-obvious naming conventions, cross-module dependencies, and undocumented rules that existed only in engineers' heads.

First attempt

Existing AI-powered operational systems could not be extended to development tasks because agents had no understanding of the proprietary config-as-code structure, causing them to produce subtly incorrect code.

Workflow diagram · grounded in source
1
Explorer agents map codebase
Ai action
Two explorer agents mapped the codebase.
▾ source quote
“Two explorer agents mapped the codebase”
2
Module analysts analyze files
Ai action
Eleven module analysts read every file and answered five key questions.
▾ source quote
“11 module analysts read every file and answered five key questions”
3
Writers generate context files
Ai action
Two writers generated context files.
▾ source quote
“Two writers generated context files”
4
Critic agents review quality
Validation
Ten or more critic passes ran three rounds of independent quality review.
▾ source quote
“10+ critic passes ran three rounds of independent quality review”
5
Fixers apply corrections
Ai action
Four fixers applied corrections.
▾ source quote
“Four fixers applied corrections”
6
Orchestration routes engineer queries
Routing
An orchestration layer auto-routes engineers to the right tool based on natural language.
▾ source quote
“we built an orchestration layer that auto-routes engineers to the right tool based on natural language. Type, "Is the pipeline healthy?" and it scans dashboards and matches against 85+ historical incident patterns. Type, "Add a new data field" and it …”
7
Automated self-refresh
Feedback loop
Automated jobs periodically validate file paths, detect coverage gaps, re-run quality critics, and auto-fix stale references.
▾ source quote
“automated jobs periodically validate file paths, detect coverage gaps, re-run quality critics, and auto-fix stale references”
Reported outcome

The pre-compute engine achieved 100% AI context coverage across 4,100+ files (up from ~5%), documented 50+ non-obvious patterns, and preliminary tests show 40% fewer AI agent tool calls per task; complex workflow guidance time dropped from ~two days to ~30 minutes.

Reported metrics
AI context coverage100%
Codebase files with AI navigation4,100+
AI agent tool calls per task40% fewer
Complex workflow guidance completion time~30 minutes
Show all 7 reported metrics
AI context coverage100%
Codebase files with AI navigation4,100+
AI agent tool calls per task40% fewer
Complex workflow guidance completion time~30 minutes
Quality critic score (after three rounds)4.20 out of 5.0
File path hallucinationszero hallucinations
Tested prompts core pass rate100%
Reported stack
large-context-window model
◆ Does this fit your context?

Compare to your context

Tell us your scale, team, and constraints. We'll show what changes at your size, what fails at your scale, and whether this case is a fit, needs adaptation, or won't scale to you. Free demo, no signup.

Compare to your context →
~30 seconds · free
Source
https://engineering.fb.com/2026/04/06/developer-tools/how-meta-used-ai-to-map-tribal-knowledge-in-large-scale-data-pipelines/
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

The pre-compute engine achieved 100% AI context coverage across 4,100+ files (up from ~5%), documented 50+ non-obvious patterns, and preliminary tests show 40% fewer AI agent tool calls per task; complex workflow guid…

What tools did this team use?

large-context-window model.

What results were reported?

AI context coverage: 100%; Codebase files with AI navigation: 4,100+; AI agent tool calls per task: 40% fewer; Complex workflow guidance completion time: ~30 minutes (source-reported, not independently verified).

What failed first in this deployment?

Existing AI-powered operational systems could not be extended to development tasks because agents had no understanding of the proprietary config-as-code structure, causing them to produce subtly incorrect code.

How is this back office ops AI workflow structured?

Explorer agents map codebase → Module analysts analyze files → Writers generate context files → Critic agents review quality → Fixers apply corrections → Orchestration routes engineer queries → Automated self-refresh.

Related back office ops cases

WHAT TO DO WITH THIS

Now compare it to your context

This case is one data point. Whether its pattern fits you depends on your volumes, your stack, and your exception load — that comparison is the step no case study can do for you.

Browse the Back office ops cases →