Minotauris 0.9.0 is a major reliability and safety release. It focuses on the parts that matter when agents run for a long time: truthful reports, steadier browser sessions, better screen control, clearer remote access, and more personal customization.
Why this release matters
Internal red-team testing showed the same thing many builders are starting to feel: collaborative desktop agents can chain planning, browser automation, screen control, code execution, and memory into surprisingly capable end-to-end workflows. In a controlled test environment, a single high-level prompt was enough for agents to progress through a complex adversarial task without step-by-step human babysitting.
That does not mean Minotauris is AGI. It does mean agent collaboration is powerful enough to deserve serious safety boundaries. Minotauris 0.9.0 tightens those boundaries while improving the everyday reliability users actually need.
We are not publishing exploit instructions, target details, or operational steps from internal tests. The point is product direction: stronger agents must be paired with clearer restrictions, better verification, and honest reporting.
Major bug fixes
- Agent reports are more reliable after long-running worker sessions.
- Supervisors and workers now have clearer upward reporting behavior.
- Message popups stay inside the app canvas instead of drifting off-screen.
- Worker file access is better constrained to the intended workspace.
- Extra purchased credits are included in the available-credit display.
Browser automation improvements
- Stealth browser sessions are cleaned up with the desktop app lifecycle.
- Personal Chrome selection is gated on whether Minotauris can actually connect.
- Browser context is clearer about whether agents are using isolated browsing or the user browser.
- Automation settings are less likely to accept a browser mode that cannot run.
Screen control improvements
- Screen actions now map image-relative coordinates into the captured monitor space more safely.
- The public UI now calls the feature Screen Control instead of exposing internal model/vendor labels.
- Completion still depends on visible final-screen verification, not only the action model claiming success.
Remote control rebuilt
The website remote-control surface has been rebuilt around the real flow: pick a computer, choose a space, choose a session, then command or steer the work. The new layout is mobile-focused, keeps schedule controls visible, and makes remote sessions easier to scan.
Making Minotauris yours
This release also continues the move toward a more personal Minotauris: customization, memory self-learning, cleaner settings, and a desktop app that feels less like a temporary tool and more like your own agent workspace.
Minotauris 0.9.0 is about making the system faster, safer, and more reliable without hiding uncertainty. Powerful agents need strong proof, visible controls, and honest limits.