Minotauris 0.6.8 is the biggest desktop experience update we have shipped so far. It touches the core UI, computer use runtime, browser automation, BYOK model support, routing architecture, memory systems, and the long-running automation loop that keeps agent work grounded over time.
Biggest desktop UI redesign so far
The desktop app now has a broader redesign across the active work surface. The new inspector makes live runs easier to understand, Assembly is a first-class inspectable surface, and the canvas is tuned for longer sessions where users need to scan decisions, browser evidence, terminal output, memory, todos, and agent state without losing context.
The Assembly node has been rebuilt around the actual shape of group decision-making: members, discussion, ideas, votes, elapsed time, and verdict state are visible in a cleaner layout. The inspector now understands Assembly directly, showing phase, vote tally, proposals, recent discussion, and member reasoning.
Next generation computer use
This release introduces Minotauris LiveUI Fabric, our next-generation computer use layer. LiveUI Fabric is built to make desktop automation more efficient, faster, and more reliable by giving agents stronger semantic state, stable targets, and a better action loop than raw coordinate driving alone.
LiveUI Fabric is still marked experimental because desktop automation is a hard surface and some apps may still break. The important change is architectural: agents now have a clearer path toward dependable local app control while still reporting failures honestly and falling back when needed.
Better browser automation and visual evidence
Browser automation is stronger in headless sandbox sessions. Agents can navigate, inspect pages through CDP, and capture screenshots when the selected model supports image input. That means browser work can stay controlled and isolated while still giving vision-capable models the visual context they need.
The steering and comment inputs now support local image context for compatible agents. Pasted images remain local data URLs until the user intentionally sends them into the agent context.
BYOK, Azure, and smarter model routing
Minotauris 0.6.8 expands BYOK support with Azure OpenAI compatibility and improves model routing around real agent roles. The new Auto mode lets higher-level agents choose models for lower-level agents: Leaders can choose models for Managers and Workers, and Managers can choose models for their Workers.
That hierarchy-aware routing makes workflows more efficient and more cost aware. Expensive reasoning can stay where it matters, while routine browser, coding, or research work can use the best model for the route you pick.
Optimization, token efficiency, and anti-hallucination work
This update includes broad desktop optimization work and engineering architecture changes aimed at smoother workflows. The agent system now does more to preserve useful state, reduce unnecessary token use, and keep long-running automation grounded in evidence.
New anti-hallucination methods improve longer automation runs by pushing agents toward inspectable evidence, explicit verification, persisted memory, and clearer reporting. The result is a system that is better at continuing work without drifting away from what actually happened.
Self-evolving system and efficient memory dreams
Minotauris is also moving deeper into self-improving local-first behavior. The self-evolving system and memory dreams are designed to learn from repeated work without wasting token budget. The goal is simple: remember durable patterns, surface useful knowledge, and avoid turning memory into noise.
Together, these changes move Minotauris closer to a real agent operating environment: a desktop app where agents can see, decide, remember, route work, control tools, and keep users oriented through complex automation.