Sandboxed code steps and the Agent Harness

Workflows gained a new Code activity: sandboxed Python as a step, running on Docker or Kubernetes. And every agent now has a home outside a workflow too — standalone agents and an Agent Library run agents directly with persisted history, backed by a new Agent Harness: an Artifacts view, an action ledger, and a Blackboard store for what agents remember across steps.

Also in this release:

• Knowledge base overhaul: SharePoint folder sync, URL and crawl ingestion, and a rebuilt PDF/Office extraction pipeline with per-page routing
• Granular LLM control: per-user and per-model token quotas, agent- and node-level model pinning, and per-message overrides
• Files in chat: ChatGPT-style upload and display, including files an agent produces mid-run
• Whitelabel branding, project-specific agents and skills, canvas undo/redo, and multiple database connections per workspace

We also closed a critical VAPT finding (public Prometheus exposure, CVSS 9.8), fixed IDOR and tenant-isolation gaps across file and vector-query paths, and added a decompression-bomb guard for uploaded Office files.