Capability-Based Task Routing
Voss Creative Studio ·
The Challenge
Voss Creative Studio ran a pipeline of AI agents for campaign production: copywriting, image generation, translation, brand-compliance review, and layout composition. A central dispatcher mapped each incoming brief to a specific agent by name. When the translation agent went down during a product launch, briefs that required translation queued indefinitely. The team wrote failover rules that routed translation tasks to a secondary instance, then wrote more rules when the secondary also needed a fallback. Every new agent or capability change meant updating the dispatcher's routing table and its growing tree of exception handlers.
The dispatcher encoded two concerns that did not belong together: what work needs doing and who can do it. Changing either required changing both.
The Solution
Voss replaced the dispatcher with Semantik. The system now has two simple conventions: tasks describe what they need, and agents describe what they can do.
Tasks as Messages, Agents as Subscribers
A project manager publishes a campaign brief into a namespace. The message carries a semantic description of the work: "translate Danish marketing copy to Finnish preserving brand voice." No agent name appears in the message. Each agent registers a SemQL subscription describing its capability. The translation agent subscribes to work involving natural-language translation between Nordic languages. The copywriting agent subscribes to original content generation. Semantik's semantic routing evaluates the brief against all active subscriptions and delivers it to every agent whose capability matches.
Subscription Overlap as Failover
Two translation agents maintain overlapping subscriptions. One covers all Nordic language pairs; the other covers Finnish and Swedish specifically with domain expertise in financial terminology. Under normal operation both receive matching briefs. When the general translation agent goes offline, its subscription disappears. Briefs that previously matched both agents now route exclusively to the specialist. No failover rule was written. The subscription overlap is the failover. When the general agent returns, its subscription reappears and load distribution resumes.
Subscriber-Governed Routing for Evolving Capabilities
When the studio added a new agent for video thumbnail generation, the agent registered its subscription and immediately began receiving matching briefs. No existing agent was modified. When the copywriting agent expanded its capability to handle tagline generation, it widened its subscription. The change took effect on the next brief without a deployment.
Results
The studio eliminated all hand-written routing and failover logic from its pipeline. Agent fleet changes — additions, removals, capability updates — take effect through subscription changes alone, with no redeployment of existing agents. During a regional campaign launch, the general translation agent went offline for twelve minutes. The specialist agent absorbed the workload automatically. The project manager was not paged and noticed only from the post-incident review that a failover had occurred.