Why
-Graph
AI agents ship the change. FalkorDB remembers why it broke. The next PR is checked before it merges.
Overview
Press Next to follow one change from a Band agent to a blocked PR.
Back
Next
Band · AI engineering team
room why-graph-demo · 5 Claude Code agents
wg-coordinator
assigns, merges
wg-engineer
writes the fix
wg-reviewer
checks with MCP
wg-analyst
asks the brain
wg-chaos
makes the mistake
GitHub · astro-shop
config/services.yaml
PR → review → merge to main
payment timeout 800 → 200
astro_sim
watches main, diffs the config
deploys · telemetry · alerts
runtime simulated from latency_ms
Why-Graph API
FastAPI · grounded agents · MCP server
On every incident, in this order
Triage
root alert, blast radius
Memory
seen this shape before?
Cause
guilty commit, as a path
Comms
status update, pages owners
Every claim cites a Cypher query and
the exact nodes; it is re-run to verify.
Claude writes the words. The graph decides.
FalkorDB
one graph per tenant
Episodic
incidents, findings
Semantic
lessons, vectors
Procedural
detectors = Cypher
services · people · PRs
commits · config · runbooks
Console · why-graph.vercel.app
Incident (live agents, evidence path, Run live) · PR Guard · Memory · Brain · Adopt it
React + @falkordb/canvas, streamed over SSE
opens PR
on merge
deploys
telemetry
alerts
Cypher
live stream · confirm the fix
MCP · check_change_risk · incident_report · why_is · who_knows
1
Agent ships a bad timeout
PR #1, merged
2
Checkout fails, alerts fire
INC-1031 opens
3
Agents prove the cause
commit → config → payment p99
4
Human confirms, graph learns
lesson L-1 + detector D-1
5
Next PR is blocked
safe value ≥ 453 ms