Skip to content
Agent Fleet

v0.18.0

A session can now start sessions of its own and look after them, models for the self-hosted engines are found and taken in without leaving the Console, image generation can run on the deployment’s own ComfyUI, and a deployment can be given a colour and a name so several environments are told apart at a glance.

New

Section titled “New”
  • A session can start sessions of its own and look after them. It hands a task to a child, lists the ones it started, reads their output, and stops or resumes them — it cannot instruct, answer for or delete any session. Off by default, in Settings > Agents, where the number of children one session may have at a time is set as well (1–6, default 3).
  • Every session can see the fleet it is part of. What used to be an opt-in is now on everywhere, including workspaces that had turned it off, and the setting is gone. Sessions can also edit a memo now, not only add one.
  • Related sessions and worktrees are nested in the left pane. A worktree sits under the session that made it, and one family shares one colour of vertical line. A session you launch from a handoff proposal now joins the family it came from.
  • A row holding a handoff nobody launched says so, instead of looking like a session with nothing left to do.
  • A deployment can have its own colour and name. Choose them in Settings > Admin > Appearance and the favicon, the home-screen icon, the login screen and the app name follow, with no redeploy — so the same image can serve dev and production and still be told apart.
  • Models are searched from the Console. Hugging Face and Civitai are searched from the ingest screen with the licence, size and context length on each row, and the catalogue can now be browsed on a deployment running no engines at all — which is exactly the deployment deciding whether to stand them up.
  • A Hugging Face token is entered in the Console, so gated repositories (all of SD3.5 and FLUX.1) no longer take a CloudFormation run to take in.
  • Image generation can run on the deployment’s own ComfyUI, covering SDXL, SD3.5, FLUX.1, FLUX.2 klein and Z-Image. A request can name the model it wants, and is told which one is already loaded.
  • The GPU an engine buys is chosen in the Console. The operator declares a ladder of instance classes for the deployment and an administrator picks a rung, so a model that needs more VRAM no longer means a stack update.
  • Memo categories have a ＋ that adds straight into that category, carrying a half-written draft and any uploaded image with it.

Improved

Section titled “Improved”
  • An engine is usable sooner. It wakes with only the model that was asked for and fetches the rest behind you, and swapping between models is faster.

Fixed

Section titled “Fixed”
  • Generated images did not reach the conversation. The card was swallowed by the collapsed work process whenever anything ran after it, so the images read as never made.
  • opencode’s launch menu offered only the default model and said to check your connection and plan, which had nothing to do with it.
  • A session told to stop when it was done stopped while a background command was still running. It now waits for it, as the idle timeout already did.
  • Asking a session to add a memo could file it in the agent’s own memory instead of your memo queue.

Upgrade notes

Section titled “Upgrade notes”
  • ECS deployments running the inference engines: delete the LlmModelFile and ImageModelFile lines from params/60-engines before updating. They are no longer in the template, and an update that still passes one stops with Parameters: [ImageModelFile] do not exist in the template. Nothing is lost — models are chosen from the Console.
  • The Control Plane restarts once on this update, even if you changed nothing else. Its task definition now carries X86_64 / ARM64 where it carried the lowercase spelling, so it gains a revision. The changeover has no downtime, and CpArch is written as before (x86_64 / arm64).
  • The two new engine capabilities are opt-in and set in the same file: ImageEngine=comfy for ComfyUI, and LlmInstanceClasses / ImageInstanceClasses for the GPU ladder. Leave them out and nothing changes. deploy/aws/ecs/cfn/PARAMETERS-60-engines.md has the values.

Agent CLI pins: Claude Code 2.1.267, Codex 0.154.0. The others are unchanged from 0.17.0.