Nexus is an agent operations platform: define, configure, install, operate, and monitor AI agents that perform work through external task systems and controlled execution nodes. It turns agents from ad hoc scripts into managed operational assets.
“Nexus-11 — a new generation. Its planning brain is called Deckard.”
Each assigned task is analyzed, decomposed into dependency-ordered subtasks, matched to a reasoning level, executed by a worker agent on a real repository checkout, then validated: pass, retry, ask, or block.
Execution-node registry with health, capacity, and labels; SSH-driven remote install; enrollment tokens and heartbeats; bounded concurrency with resource caps on every worker.
Named routes to local inference (Simurg), online APIs, or CLIs — with health checks, quota classes, and fallbacks. Local-first while local is healthy; fail closed when policy demands it.
Every run records its source, agent, profile, node, artifacts, usage, and verdict in an append-only event stream. Nexus records what happened — always.
Central stores credential references, never values; external task systems see a safe catalog projection, never command templates or keys.
Crash-safe by design: durable intents are replayed on restart, so a run either completes or leaves an honest record of where it stopped.
Architecture first: the control plane is actively hardening toward production, and the copy here stays architectural on purpose.
A private briefing on agent operations: definitions, execution nodes, model routes, and the run ledger your auditors will ask for.