v0.17.0

An ECS deployment can now run inference on its own GPU, on an instance that only exists while somebody is using it: a model that appears in opencode’s launch menu, and image generation that needs no external account. Sessions can also be told to stop themselves once the work is done, and image generation gained a second provider.

New

The fleet can run its own inference engines. A GPU instance is bought when something asks and stops itself once nobody is using it: the llm role puts llamacpp/<model> in opencode’s launch menu, and the image role serves generate_image from the deployment itself. Administrators choose disabled, always on or on demand in Settings > Admin > Inference engines. ECS deployments only, opt-in, and not cheap while it is up — see the Upgrade notes.
Models are added and swapped from the Console. “Take one in from Hugging Face” pulls a model into the deployment’s own bucket, licence acceptance and gated repositories included; a file already in the bucket can be registered by hand. Each row can be enabled, chosen as the one loaded at start, or forgotten — with or without deleting the bytes — and none of it needs a CloudFormation run.
The engines panel says what the instance is doing: when it started, when it will stop itself, what has been asked of it recently, which model is in VRAM, and 14 days of occupancy.
A session can be told to stop when it is done. It finishes the turn it is running and then halts, so the conversation and the working copy stay and it can be resumed later. The session’s ⋯ menu offers it, and so can the agent itself and the operator.
A scheduled run can stop its session afterwards. A workspace woken at 3am no longer stays up until the idle timeout.
A session can look at the fleet it is part of — whether a session it handed work to is still going, how full its own context is — and leave what it noticed in your memo queue. It gains nothing that acts: instructing, answering, approving, starting and stopping a session stay with you and the operator, and the text of a question or plan another session is waiting on is not passed to it. Off by default, in Settings > Agents.
Image generation has a second provider. Antigravity’s built-in generation (Gemini, the “Nano Banana” models) spends a different plan from Codex’s, honours the aspect ratio you ask for, and is now tried first. A request can name the provider it wants, an automatic one moves on to the next when a plan is out of quota, and the tool now says which service it went to.
Claude’s bundled skills are in the mirror’s skill picker. Your own skills come first; the CLI’s fold into one line that opens with //, with a second press of the trigger, or with ↓ then Enter.
The shared-cost card says what the cost was for. Inference engines and the speech engine are their own lines instead of being lumped into EC2 and ECS.

Improved

Tabs of sessions that went away close themselves. A session archived from the cleanup screen, from another device, by the operator or by an agent left a tab that said “no session” and could never reconnect.
After you re-authenticate, the card says so and the session carries on by itself, instead of showing “Re-authenticate” until you typed “continue” yourself.
Making a worktree of a repository with large submodules no longer downloads them again — they are seeded from the clone that is already on disk.
Antigravity can be used on hosts whose kernel has withdrawn RDRAND, where it was hidden from the launch menu entirely.
Sessions spend less of their context on the descriptions of Agent Fleet’s own tools.

Fixed

In a long conversation the work process opened by itself and the reading position jumped, both while scrolling and when older messages were loaded.
A half-written free-text answer to a question card was lost when the session stopped or resumed and the card was replaced by the other kind.
The file viewer’s font-size setting did not reach Markdown, and dragging the minimap selected the text it passed over.
A path in the mirror could be swallowed by a commit link and stop opening — the paths of generated images in particular.
A notification with no session to open said the session could not be found.
A launch that changed the slot type said it would never start when it was about to start half a minute later.
Codex’s usage chip ignored image generation, so the quota it was spending did not move.
opencode’s launch menu listed amazon-bedrock/* models that no workspace can reach.
On a phone the workspace bar had no label at all, where it now says “WS”.

Upgrade notes

Nothing to do unless you want the self-hosted engines. They are opt-in and ECS only: update 20-platform and 30-ingress, then deploy the new cfn/60-engines.yaml stack. deploy/aws/ecs/README.md has the steps, including the G-family vCPU quota increase — a new account has none, and the request takes hours to days to be granted. A GPU instance costs $1.26 an hour while it is up, which is why “on demand” is the mode to leave it in.

The cost screen makes one more Cost Explorer request per refresh (about $1.2 a month). Cost tags do not apply retroactively, so the new per-role breakdown starts from the day you deploy.

Agent CLI pins: Claude Code 2.1.266, OpenCode 1.18.30, Antigravity 1.1.28, Cursor 2026.09.08, Kiro 2.21.2. Codex and Copilot are unchanged from 0.16.0.