Prime Agent

Install it on macOS or Linux with a single script that downloads a versioned release, verifies its SHA-256 checksum and prepares the Python runtime. Change into the project you want worked on and run prime-agent. On first launch, /login picks a subscription or API-key provider - the agent brings no inference of its own.

From there it behaves less like a chat and more like a process you manage. Sessions are daemon-backed, so they keep running when the terminal disconnects and can be reattached later with prime-agent attach or browsed with prime-agent agents. /goal keeps an objective and its progress alive across turns until you clear it. /heartbeat and prime-agent schedule re-enter a session periodically or at a set time. /autonomous continues within configured turn, token and time budgets and can run quality gates you define, and the documentation is careful to note that a passed gate checks only what that gate verifies and that hitting a limit does not imply the task succeeded.

• •Programmatic Everything - A persistent Python REPL is the primary model tool, so file edits, shell commands, tool use and context management are written as code rather than selected from a fixed tool list
• •Built-In Subagents - rlm(...) spawns real child agents for parallel or background work and returns their results programmatically, and running agents can message and steer one another directly
• •A Harness That Improves Itself - /refine persists focused, reviewable lessons as supplemental prompts, memories, skill descriptions or subagent specs, with refinement history and snapshot rollback, and without rewriting the base prompt
• •Executable Skills - Skills are importable Python packages, and a built-in skill creator turns a recurring workflow into a project or personal skill
• •Daemon-Backed Continuity - Active sessions, REPL state, schedules and subagents survive terminal disconnects and can be reattached
• •Long-Horizon Controls - Automatic compaction, persistent goals, heartbeats, schedules and a bounded autonomous mode with turn, token and time budgets plus user-defined quality gates
• •Headless Modes - JSON mode and RPC mode for automation and integration into other systems

Researchers running long evaluations, and engineers with tasks measured in hours rather than prompts - migrations, large refactors, work that needs to survive a closed laptop. If you already think in Python, the RLM model is the appeal: composing subagents and manipulating context as values is far more expressive than a fixed tool schema, and skills as importable packages means your workflows are real code you can test.

It is the wrong tool for several people. If you want autocomplete or a quick in-editor edit, this is much heavier than you need. If you cannot run it against a disposable checkout or inside a container, the project's own warning about the absence of a sandbox should stop you. And if you are choosing your first terminal coding agent, something more conventional will be a gentler start - the value here is concentrated in the long-running and self-refining behaviour, which only pays off on work that actually runs long.