| Primary unit of work | A conversation in the Antigravity app or command line, organized into projects. Agents can work in the background, and a parent agent can start several subagents at once. | A mission: one goal with its conversation, the agents working on it, their approvals, and its record. It persists across sessions and goes idle rather than closing. |
|---|
| Execution environment | A desktop app, the agy command line, extensions for five IDEs, and a Python SDK. Remote Control drives desktop and command-line sessions on your machines from a web browser. | Agents run on your Mac by default. On a paid plan you can start or move a mission to Skyflo Cloud, a machine Skyflo runs for that mission, which keeps working while your Mac sleeps. Only Codex and Skyflo's own agent run there; Claude Code stays on your Mac. A mission started there from a GitHub repository opens a pull request. |
|---|
| Agent and model choice | Google's own agent, on Gemini models such as Gemini 3.8 Flash, with Gemini 4 Argon on Google AI Pro and Ultra. Ultra adds third-party models; the model list includes Claude Sonnet 4.6 and Claude Opus 4.6. You choose a thinking level per conversation. | Skyflo's own agent plus seven external agents: Codex, Claude Code, Cursor Agent, Antigravity, Grok CLI, OpenCode, and Pi, and one custom ACP agent added through a manifest file. External agents use their own sign-in and models; Skyflo's own agent uses included models on paid plans or your Moonshot or Z.ai key. Orchestrate runs several on one mission and can pick the lead. |
|---|
| Cross-repository coordination | A project is a configuration of one or more folders, for example a frontend and a backend repository, with its own settings and permissions. A subagent can work in the parent's folder or in its own Git worktree. | A workspace is an ordered set of up to eight registered Git checkouts, and one mission can span them. In Skyflo Cloud a mission spans one to five. Each writing specialist works in its own worktree, so edits stay apart while the work runs; integrating them is still a change you review. |
|---|
| Approval and mutation boundaries | A permission engine checks each sensitive action against Deny, Ask, and Allow lists layered on a preset, and Deny takes precedence. The updated engine covers macOS and Linux; Windows keeps the previous system. The command line also has a sandbox mode and a flag that skips permissions. | One access mode governs every agent on a mission, Skyflo's own and the external ones. Skyflo answers each agent's permission requests by that mode and records the decision. In Full access, the default for a new Auto mission, prompts are off and folder confinement is lifted. Orchestrate starts at Ask for approval, and Plan mode waits for you to approve the plan. |
|---|
| Independent review | Teamwork agent teams, in preview on paid plans, run independent verification checks at every milestone, with a Success Auditor before the work wraps up. | In Orchestrate missions, Skyflo adds a separate reviewer by strategy and risk, and Skyflo's own agent can start one. The reviewer can read, search, ask, and report, but cannot edit files. Not every mission is reviewed, and a review is a second opinion, not a guarantee. |
|---|
| Browser, terminal, and automation evidence | A browser subagent tests in sandboxed browsers through the /browser command, scheduled tasks run on a cron schedule in the background, and a terminal sandbox isolates shell commands inside OS-level boundaries. | Code, browser, terminal, and automations share one workspace, where you and Skyflo's own agent work in the same browser and terminal. Each mission keeps an append-only record on your Mac of approvals, Skyflo tool calls, including external agents' calls through Skyflo's tools, and review findings, every event hash-linked to the one before it. |
|---|
| Cross-mission memory, provenance, and user control | Rules and agent skills provide persistent instructions across conversations. Subagents start with a clean context rather than the parent's history. | Personal memory stored on your Mac, learned from finished missions with a link to its source. Skyflo keeps a memory when the evidence supports it; otherwise it waits for you to keep, correct, or dismiss it, and you can remove any memory later. A workflow verified in two repositories can become a Skill once you or an independent agent approves it. Reverting a Skill or a routing change alters what Skyflo does next, not past work. |
|---|
