On-call agents that remember every incident.
Playbooks live on the service graph. Agents recall them by walking dependencies, check them against live evidence, and page one engineer, once.
Built on FalkorDB, for on-call teams and their agents.
Remember
A confirmed incident becomes a playbook, pinned to the service that failed.
Recall
A new alert starts a walk along service dependencies to nearby playbooks.
Check
Each playbook is tested against live evidence before it is trusted.
Act
Leased tasks, one diagnosis, one confirmed page to the on-call engineer.
Learn
Trust rises when a playbook helped and falls when it misled.
Memory is found by walking the dependency graph.
When payments-gw alerts, the agent walks to postgres-payments, a datastore of the same kind as postgres-orders, where a playbook was learned. A text search for “slow responses” would not have found it.
Nothing is trusted until live evidence agrees.
Every playbook carries checks. The agent reads the observed value for each one before it trusts the playbook. In s04 every check holds. In s05 the same playbook is recalled, but checkout-api was just deployed, so it is ruled out and its trust is left alone.
Kill an agent. Its work carries on.
Agents claim tasks with a lease and write a checkpoint after every step. When an investigator is killed, its lease drains over six seconds. The other claims the task, the epoch moves from e1 to e2, and the work resumes at the next step, not from scratch. The dead one can never write again.
One page, the way the engineer asked.
Trust is earned per playbook, and lost the same way.
Engineer preferences are versioned facts. What the engineer approved supersedes what the agent assumed. Both stay on record.
Memory survives a database restart.
Restart FalkorDB mid-demo. Mission control dims, reconnects, and reads everything back from the graph.
One graph. Every step is a query you can read.
Services, playbooks, incidents, leases, checkpoints and conversations all live in FalkorDB. Every panel in mission control has a query button that shows the Cypher behind it. Every step is stored, so any incident can be replayed.
Watch the agents work an incident.
Five scenarios. Kill an agent, restart the database, or answer the war room yourself.
Enter mission control →