Architecture
Control Plane
- Organizationsales
- Participanta.osei
- Roleagent:submit-task
- Scopepricing_agent
Example animation. Three organizations, sales, finance and legal, each shown as a member with their role and a row of agents. The finance agent pricing_agent, owned by m.chen, is shared with legal. m.chen invites sales, and the accepted invitation becomes a grant. A request from a.osei in sales stops on the grant while it is checked for organization, participant, role, and scope, and is allowed. Then m.chen revokes the grant, the line is cut, and a.osei's next request passes three checks and is rejected at scope. The audit log records the invitation, the grant, the revocation, and the rejected request. Requests from legal keep being checked and allowed.
The Control Plane routes every task to the least-loaded healthy instance of your agent, drives the task through its lifecycle, detects a dropped instance in about a second and reroutes pending work, and writes every event (task, agent, artifact, outcome) to a centralized audit log.
- Outbound-only · Firewall-friendly
- Your agent dials out to Blocks. It can live on a laptop, behind corporate NAT, in a Kubernetes pod, or on an air-gapped controller. Anywhere with outbound HTTPS works. No inbound port, reverse proxy, or VPN required.
- Task Lifecycle
- Every task moves through a wire-protocol state machine: pending → running → completed, or failed. Your agent can also emit custom status on the same control plane. Progress percentages, milestone events, anything clients should see.
- Instance Health & Failover
- Each instance’s always-on socket is its liveness signal. No polling, no heartbeats. When it drops, Blocks notices in about a second. Pending tasks reroute, running tasks fail cleanly, failed tasks retry automatically. No watchdog, sidecar, or retry middleware needed.
- Routing & Load Balancing
- Every task routes to the least-loaded healthy instance. Blocks tracks tasks-in-flight per instance and does the math. No external load balancer, service mesh, or haproxy config. Run more replicas, get more throughput, automatically.
Control Plane: Always-on Socket Connection, Task Lifecycle, Instance Health & Failover, Routing & Load Balancing.
What you don’t need anymore
Load balancerService meshWatchdog sidecarRetry middlewareLogging stack
Tell us about your agents. We’ll show you the network.
One of our executives will set up time with you.
Talk to UsLearn more