An agent that treats prompts as variables and subagents as function calls

What it does

Prime Agent is a coding and research assistant designed for long-running tasks. It wraps a persistent IPython environment in a daemon-backed session that survives terminal disconnects, allowing goals, schedules, and subagents to keep working while you are away. The agent executes model-generated Python and shell commands in your local project directory, storing supplemental prompts, memories, and reusable skills as durable state it can refine over time.

The interesting bit

The Recursive Language Model abstraction treats context as variables and subagents as programmatic function calls inside a persistent REPL, which means parallel or background work is spawned through rlm(...) rather than orchestrated through chat messages. This is paired with a Continual Harness that stores working memory and skill descriptions outside the immutable base system prompt, letting the agent improve its own supplemental state through small, evidence-backed updates without losing rollback history.

Key highlights

• Daemon-backed sessions keep IPython state, subagents, and schedules alive after terminal detach, with reattach support
• Built-in subagent spawning lets running agents discover one another and exchange messages directly without routing through the user
• A refinement mechanism reviews trajectories and updates supplemental harness state (memories, prompts, skill specs) with recorded snapshots for rollback
• Skills are importable Python packages, and the agent can turn recurring workflows into reusable, executable capabilities
• Bounded autonomous mode operates within configured turn, token, and time budgets, with user-defined quality gates

Caveats

• The README warns that model-generated code runs with your user permissions and the worker/kernel isolation is for recovery, not security; untrusted code needs an external sandbox
• It is built on top of the external pi TUI framework, so a meaningful portion of the stack lives outside this repository
• The project requires a subscription or API-key provider to function, so it is not a bring-your-own-model fully offline tool by default

Verdict

Worth a look if you want an agent that treats long-running research or coding tasks like durable background jobs rather than chat threads. Skip it if you need a fully offline, sandboxed coding assistant or prefer conversational UIs over programmatic REPL control.