Kiva engineer details running 6–10 AI agents in Orca

The write-up is less a product review than a field report on what changes when a single developer tries to keep six to ten AI agents working at once. The engineer moved to Orca from cmux, noting that the basics of automatic worktree creation and tabbed session switching did not differ much; the divergence appeared only after the number of parallel sessions grew. Beyond six, the demands shift to wanting work to continue while away from the desk and to keeping per-worktree dev servers from breaking each other. Orca's answer to the first is remote access through Tailscale, which the engineer uses to let a sales colleague ask questions of the repository from a phone. The second is framed as something refined through real use rather than a built-in guarantee. The release cadence data supports the picture of a fast-moving tool: 386 official desktop releases in roughly eight months, with 91 days seeing two or more releases and 49 days seeing three or more. Orca's chat interface, first shipped as Native chat in v1.4.144 on 2026-07-17, saw near-daily fixes over about two months for issues such as IME input, Windows path display, mobile sending, and attachments over SSH. The practical lessons center on resources that are shared by default on a single machine. The engineer describes port collisions, cookie collisions across localhost ports, a shared worker registry, and a shared built-in browser whose active tab can be stolen by another session. Nested worktrees created by sub-agents are another case: they live outside Orca's view and can accumulate, with the engineer counting 13 of 25 worktrees on 2026-11-10 as sub-agent creations totaling 42GB. The suggested takeaway is to inventory everything that exists once per machine before scaling up parallel development.