Parallel agents are only useful if you can see what each one is doing and if your quota survives them. This week’s releases address both: swarms bring worker agents and a live board, and Auto becomes the default Agent Mode, so unattended runs now start with different permission behavior than before. A scripted run_workflow tool, new Gemini 4 models and richer hook input round out the batch.
Week 46 (November 10-14, 2026) covers versions 1.5.4 to 1.5.9: six releases and 95 changelog entries. Swarms are the main topic, followed by the changes that touch everyday sessions, then a set of technical notes for integrations and admins.
Swarms: several agents on one task, and a board to watch them
What it is. Swarms arrive as a preview in 1.5.5. You start one with /swarm <task>, and the agent becomes a coordinator that adds worker agents. Each worker runs in its own conversation, works on the task in parallel and collaborates with the others through a shared message board.
How workers are set up. A worker can be forked from an existing member’s conversation, so it starts with that member’s context. Each worker can also run on any model in the /model picker. Workers start from your uncommitted work (staged, unstaged and untracked), and non-Git workspaces can get an isolated copy.
What happens when things go wrong. Workers pause instead of retrying when their model runs out of quota. A monitor worker alerts the coordinator when the whole swarm stalls. (Whether the monitor catches every stall in practice is something you will only learn by running one.)
The board. /swarm board, or alt+s (rebindable as swarm.board), opens a full-screen live view of the current swarm with three panes:
Rosterlists the members, each with its state and model.Postsis the feed of what the workers publish.Messagerenders the selected post as Markdown.
Navigation is keyboard-driven. tab and shift+tab move between panes, and z zooms the focused pane. enter shows or hides the message, and [ and ] step through posts. p filters the feed to one member’s posts and the posts that mention them. o opens a member’s subagent details, f follows new posts and esc backs out.
The tray row. While a swarm runs, a ● Swarm(<title>) row appears in the active-items tray below the prompt. It shows the latest post and the post count, then Done when the swarm finishes. Press Enter on the row to open the board, or alt+x (swarm.dismiss) to dismiss it.
The cost. Every worker is a separate agent that uses your quota. That is why /swarm is labeled (Preview & Quota Heavy) in the / menu. Treat the label literally: start with a task that justifies several agents, and watch the Roster to see which models are burning through quota.
Changes for everyday work
Auto is the default Agent Mode. Sessions with no saved mode, including headless -p runs, now start in Auto wherever Auto mode is available, and in Default elsewhere (1.5.7). To keep the previous behavior, choose Default for Agent Mode in /config, which saves "agentMode": "" to settings.json. If you run -p in scripts, check this first, because those runs now start in a different mode than before.
strict still wins. A Tool Permission setting of strict now takes precedence over Auto and accept-edits modes, so file edits still ask for approval under strict. That is the setting to use if you want Auto for commands but a human for edits.
Auto in editors. agy acp sessions can now be switched to auto wherever the CLI itself offers Auto mode (1.5.4). In that mode the agent approves edits inside the workspace and runs commands in the terminal sandbox. It still asks before commands that would leave the sandbox, unless Auto mode’s review clears them.
Scripted fan-out with run_workflow. The agent can now run a Python workflow script that fans work out to many subagents, in parallel or in stages, and then combines their results (1.5.9). It uses the Python bundled with the CLI, so you do not need Python installed. The tool is available to everyone. Unlike a swarm, the fan-out is defined by a script, so the structure of the work is fixed up front.
Gemini 4 models in /model. What you see depends on how you sign in:
- With Application Default Credentials (
AGY_ADC_AUTH),/modeland--modeloffer Gemini 4 Carbon (Low, Medium, High) from 1.5.5 and Gemini 4 Hydrogen (Low, Medium, High, XHigh) from 1.5.8. - With a
GEMINI_API_KEY, you get both Carbon and Hydrogen (1.5.8). - Hydrogen is also supported through an enterprise model gateway.
Click to expand in full screen mode. In the full screen (altscreen) rendering mode, clicking a tool group, a thought, a prompt’s attached images or multi-line command output expands or collapses just that item (1.5.8). You no longer need Ctrl+O for this. The default Rendering Mode is unchanged, so you need to be in full screen mode to use it.
Richer hook input. The JSON that hooks receive now includes hookEventName, cwd and toolCall.id (1.5.4). One hook command can handle several events, knows the agent’s working directory, and can match a PostToolUse call to its PreToolUse call.
Technical notes
Hooks. Stop hooks get a continuation_active field (1.5.6). A hook such as ‘run the tests before finishing’ can use it to avoid looping.
Auto mode details.
- Pay-as-you-go Google Cloud accounts signed in with ADC now use the standard pay-as-you-go permission settings, which makes Auto mode available to them (1.5.6).
- Auto mode’s review of out-of-sandbox commands became faster across 1.5.4, 1.5.6 and 1.5.9. It also asks less often on safe reads such as GitHub API reads, single-file downloads and public SSH files.
Editors and remote sessions.
agy acpsessions now send context window usage updates, so editors such as JetBrains AI Chat can show how full the context is (1.5.8).- The first message in
--remote-agentsessions no longer waits for the remote environment to start or resume (1.5.8).
Session handling and rules.
/resume <id>accepts any unique prefix, such as the short ID in the status line. The picker also no longer stalls on ‘Loading conversations…’ (1.5.9)./learnnow checks rule size limits before updating (1.5.8). It proposes condensing or splitting instead of growingGEMINI.mdorAGENTS.mduntil they get truncated.
Plugins. agy plugin import and install now convert Claude Code commands to skills, convert supported hooks, honor the manifest skills field and install atomically (1.5.5, 1.5.6).
Enterprise and admin.
- Enterprise model gateway URLs must now use
https://unless the gateway is local (localhost,127.0.0.0/8,::1). Plainhttp://remote gateways are rejected (1.5.5). - New admin settings are
pluginControls(pluginsDisabled,marketplaceDisabled),gateway.routeWebFetchandfeatureControls.remoteControlEnabled(1.5.8, 1.5.9). - Set
ANTIGRAVITY_DISABLE_EMBEDDED_RIPGREP=trueto stop the CLI from extracting its bundled ripgrep, for example when endpoint security blocks it (1.5.5).
What to take away
If you work in the terminal. Check which Agent Mode you are in, since Auto is now the default where available. If you prefer the old behavior, set Agent Mode to Default in /config. Try full screen mode and click items to expand them. If you want to see which Gemini 4 models you can use, open /model after signing in and compare the list with your sign-in method.
If you run parallel sessions. Start a small /swarm on a task that splits cleanly, open the board with alt+s and learn the keys (p to filter by member, f to follow). Keep an eye on quota, because every worker draws on it and the feature is labeled preview. For work whose structure you already know, compare it with run_workflow, which fans out from a script. Review your -p scripts for the new Auto default as well.
If you administer integrations or company deployments. Check that your enterprise model gateway uses https:// unless it is local, because remote http:// gateways are now rejected. Look at the new admin settings (pluginControls, gateway.routeWebFetch, featureControls.remoteControlEnabled) and decide which you need. Where endpoint security blocks the bundled ripgrep, ANTIGRAVITY_DISABLE_EMBEDDED_RIPGREP=true is the switch. If you rely on strict for Tool Permission, note that it now overrides Auto for file edits.
Sources: Antigravity CLI changelog, versions 1.5.4 to 1.5.9 (November 10-14, 2026), and the Antigravity CLI documentation. Swarms are a preview feature, so their availability and behavior may change.