v4.0.0
Version v4.0.0 — 2026-09-03
Changelog
New
- Introduced Situation Handlers — agents can react when specific events occur or when the shared runtime state matches defined conditions.
- Introduced an explicit Agent Loop — inference, tool calls, model responses, and idle states are now represented as separate steps that can be observed and controlled.
- Added Interception Handlers — developers, users, or other agents can pause, continue, or stop an agent loop while it is running.
- Added Shared Runtime State — participants can coordinate through a common state and implement mechanisms such as locks, semaphores, ownership, and handoffs.
- Added Participant Manifests — participants can describe their identity and capabilities when joining an agentic environment.
- Added Mozaik Cloud integration — runtime events can be sent to Mozaik Cloud to monitor agent loops, inspect collaboration, and understand what is happening inside a multi-agent system.
Improved
- Introduced a more granular architecture for building concurrent AI agents.
- Separated agent behavior from the conditions that trigger it, making agents easier to adapt at runtime.
- Improved visibility and control over every step of agent execution.
- Made coordination between agents, humans, and other participants more explicit.
Contributors
Contributors
Mijura and Lotus015