The agent keeps working after your terminal leaves
Prime Agent's latest release is v0.9.8, and its sessions can keep running after your terminal disconnects. The model gets a persistent Python REPL, can spawn child agents, and can leave work behind a daemon. Saved goals, schedules, heartbeats, and automatic context compaction are meant for jobs that take longer than one sitting. That persistence is the reason to consider Prime Agent.
A session can retain Python state and subagents, then accept a reconnect through prime-agent attach. The /goal command keeps an objective active, while autonomous mode can stop on time, token, or turn limits. The README states an important limit: reaching a budget does not mean the task succeeded, and passing a quality gate proves only what that gate checked.
The Python control plane adds power and state to inspect
Our 27.5 MB checkout routes shell work, file operations, tool calls, context handling, and child-agent orchestration through code in the built-in Python environment. That gives an experienced user one programmable surface for complicated workflows. Skills are Python packages rather than prompt snippets, and /refine can save small lessons into supplemental harness state with snapshots for rollback. The base system prompt stays immutable.
You also have more state to understand. Prime Agent has a daemon, workers, a kernel, live and saved sessions, reusable skills, and agent-to-agent messages. The measured checkout held 1,416 files and about 593,768 lines of source. A developer wanting one reviewed patch may find that machinery excessive. A research loop that must survive a dropped SSH connection has a much stronger reason to accept it.
What happened when we ran it
Our measurement setup was an unprivileged lab-rust:1 container with 3 CPUs, 12 GB of RAM, and no secrets. It installed commit 419777a in 11 seconds and pulled 319 packages. The Rust build completed successfully in 268 seconds. Those results cover repository setup and compilation. We did not measure agent quality, model cost, or task completion.
Cargo test did not finish before the 900-second limit. Every completed result in our supplied summary was green: 775 tests passed and 0 failed. The final log showed a TUI group finishing 41 tests in 28.90 seconds, followed by the 2-test interactive_fork_e2e.rs binary. interactive_fork_child_mode passed, but the log did not show a result for the other test before the cap. It does not identify a failed assertion or explain the timeout.
For release planning, the suite is neither broken nor confirmed clean by our run. The 775 passes rule out a blanket failure, while the unfinished command rules out a completed gate. A team adopting Prime Agent should reproduce the full command on its own runners and choose a timeout from that result, since 900 seconds was insufficient in our container. The repository has 8 CI workflow files, but no Dockerfile and no top-level tests directory in our scan.
Generated code runs with your account's permissions
The latest release is v0.9.8, and Prime Agent can execute model-written Python and project commands with the permissions of the user who launched it. Worker and kernel processes help with recovery and lifecycle separation, but they do not create a security boundary. The README recommends a disposable clone, clean worktree, or checkpoint you can inspect and restore. Untrusted repositories, instructions, and skills belong in an external sandbox.
That warning decides whether unattended use is sensible. Autonomous mode can continue within explicit budgets and run user-defined checks, but those controls limit a workload rather than contain it. A scheduled session that wakes up 8 hours later still has the same filesystem and command authority. If your normal account can read production secrets or deploy code, give Prime Agent a narrower account or container before relying on its long-running features.
Login and Windows beta fallback remain part of setup
The release installer makes the first binary easy to obtain: Unix gets a one-command shell installer, while Windows uses PowerShell. After that, /login asks for a supported subscription or API-key provider. Windows has a qualification that is easy to miss: the installer falls back to a beta build when the stable channel has no Windows artifact, and it tells the user when that happens.
Remote terminal users have another practical wrinkle. Clipboard handling varies across tmux, OSC 52 support, and terminal settings. The README documents the outer-tmux case and notes that macOS Terminal does not support OSC 52. This unusually specific guidance also shows how much Prime Agent depends on its terminal environment. Version v0.9.8, released September 29, fixed model discovery and a stale model picker rather than changing that operating model.
October 8 activity supports an active, fast-moving verdict
GitHub recorded a push on October 8, 2026, and listed 111 open issues and pull requests together. The 30 most recently updated items included open pull requests for worker journals, OAuth credential refresh, Windows bugs, and daemon behavior. With 21,604 stars and v0.9.8 published on September 29, this is an actively maintained project. The amount of concurrent work is also a reason to pin a release instead of following the main branch.
Prime Agent earns its complexity when work must persist and delegate across reconnects. Our run proved that commit 419777a installs and builds, and it showed 775 passing tests. It did not prove that the complete suite fits inside 15 minutes. Adopt it for long work only after you have given its long-lived processes a deliberately small blast radius.