One request, a coordinated agent team
Turn on Ultra mode and the primary agent can split a complex request into two to four bounded AgentRuns, follow each subagent live, exchange input when needed, recover interrupted work, and synthesize one final result. Delegation happens from the normal composer — there is no separate task-batch form to fill out.
A live work surface that stays understandable
The conversation now presents prose, reasoning, tool activity, compaction, subagent coordination, stops, and follow-up work in one ordered timeline. A unified task panel keeps progress, outputs, active and completed subagents, sources, inputs, and context close at hand, with responsive overlay and pinned layouts.
Built for long-running work
Streaming and multi-agent lifecycle handling were strengthened across stop, queued cancellation, reconnect, replay deduplication, immediate resend, backend restarts, waiting-for-input states, and parent-child completion. Permission memory is more precisely scoped, and source evidence is validated and redacted before it is saved or rendered.
Desktop polish throughout
OpenYak now includes app-wide keyboard shortcuts, a slash-command menu, localized first-run onboarding, blocked-session attention states, and a cross-session automation inbox. The conversation and composer use a tighter semantic design system with improved accessibility and reduced-motion behavior.