Neuro-symbolic agents on FalkorDB

Rehearse the outage before it happens.

ButterflyOS forks your production graph, takes down any combination of services and people, and lets a swarm of agents compute the blast radius, find the knowledge that just walked out the door, and heal the organization, all in milliseconds.

Cypher + GraphBLAS algorithmsMCP tool executionNo Node, No Action provenance
Edge GatewayAuth EnginePaymentsToken KVLedgerArchitectSuccessorWarm replica
~30–80 ms
to fork production into a sandbox
44 → 1
Cypher queries, cold start vs. vaccinated
5 of 5
orphaned systems re-homed per scenario
0
invented nodes. Every claim is verified

Why graphs

Two nightmares. One connected question.

“What breaks if this service goes down?” and “What happens if our lead architect quits?” are traversals, not text searches. Flat retrieval can't compose the edges that matter.

Technical blast radius

The cascade nobody wrote down

Payments → Ledger → Token KV is not in any document. It exists only as the composition of two dependency edges. PageRank names the chokepoint; multi-hop Cypher gives the exact failure path.

Human bus factor

The expert who left on Friday

The one warning about shard 14 sits in a Slack thread linked to a commit, a repository and a service. Components analysis finds the orphaned subgraphs; agents rebuild the context and pick a successor.

How it works

From “what if” to a rewired organization.

01

Fork

GRAPH.COPY clones production into an isolated sandbox tenant in tens of milliseconds. Production is never touched.

02

Shockwave

Multi-hop Cypher and algo.pageRank trace the cascade. algo.betweenness and algo.WCC expose ownerless subgraphs.

03

Heal

Four agents call graph tools over MCP, reconstruct tribal knowledge, and rewire ownership with auditable edges.

04

Vaccinate

The episode is distilled into a procedural playbook. The next identical failure is recalled with a single query.

The swarm

Four agents. Every hand-off is a node in the graph.

No conclusion is accepted unless every node in its provenance path exists in FalkorDB. No Node, No Action.

1

Shockwave-Triage

Maps the failure cascade, ranks chokepoints, totals RPS at risk and breached SLAs.

2

BusFactor-Detector

Finds systems whose only expert is gone and flags code that contradicts policy.

3

Tribal-Archaeologist

Walks commits, Slack and ADRs to write a cited handover for every orphan.

4

Continuity-Healer

Scores successors in Cypher and rewires ownership, blocked by policy where needed.

Built for all three tracks

One graph, three ways to make agents useful.

Track 01

Agents that act on connected data

Seven MCP tools. Every action is chosen by graph topology and returns its Cypher, latency and provenance.

Track 02

Agent memory & coordination

Episodic hand-offs, live semantic topology and procedural playbooks, in isolated multigraph tenants.

Track 03

Company brain

People, repos, commits, decisions and chats in one graph. Ask who owns what and trace every answer to its source.

Architecture

Neuro-symbolic by construction.

The graph decides what is true. A pluggable language model only phrases it, and is rejected the moment it cites a node that doesn't exist.

Symbolic core · FalkorDB
  • OpenCypher traversals and scoring
  • algo.pageRank · algo.betweenness · algo.WCC
  • graph.copy multigraph sandboxes
  • MCP tools with provenance_node_ids
  • Episodic → procedural memory
Neural layer · pluggable
  • OpenAI, Groq or Gemini synthesis
  • Constrained to nodes the graph returned
  • Output validated before it is shown
  • Deterministic narrative as the fallback

Break production safely.

Pick a scenario, press simulate, and watch the organization heal itself.