Prime Agent is a coding and research agent for general and long-running work, built on two abstractions. The Recursive Language Model (RLM) treats context as variables and tools, including recursive subagents, as function calls inside a persistent Python REPL: file operations, shell commands, tool use, subagent spawning, and context management all happen through code rather than through a fixed tool schema. The Continual Harness stores supplemental prompts, memories, skill descriptions, and reusable subagent specifications as durable state that the agent can refine in small, evidence-backed steps.
Refinement is explicit. /refine reviews the current trajectory and can apply updates to that supplemental harness state, never to the immutable base system prompt, and recorded snapshots support rollback. Skills are importable Python packages, and a built-in skill creator turns recurring workflows into project or personal skills. The underlying approach is described in an arXiv paper, and the agent and TUI are built on top of Pi.
Sessions are daemon-backed: an agent, its REPL state, schedules, and subagents keep running when the terminal disconnects and can be reattached with prime-agent attach. Running agents can discover and message one another directly. For long tasks there is automatic compaction, /goal for persistent objectives, /heartbeat and prime-agent schedule for periodic or timed re-entry, and /autonomous for bounded runs within turn, token, and time budgets with user-defined quality gates. JSON and RPC modes support headless automation. The agent executes model-generated Python with the user's permissions and its worker and kernel processes are not a security sandbox, so untrusted work belongs in an external sandbox.
Features
- Persistent Python REPL: the built-in model tool; files, shell, tools, and subagents are driven through code
- Programmatic subagents: rlm(...) spawns child agents for parallel or background work and returns their results in code
- Continual Harness: /refine persists reviewable lessons as prompts, memories, skill descriptions, or subagent specs, with refinement history and rollback
- Executable skills: skills are Python packages; a skill creator packages recurring workflows
- Daemon-backed sessions: detach and reattach running agents, REPL state, schedules, and subagents
- Agent-to-agent messaging: running agents and retained subagents exchange messages and steer each other's work
- Long-run controls: automatic compaction, persistent goals, heartbeats, schedules, and retained subagents
- Bounded autonomous mode: turn, token, and time budgets plus user-defined quality gates
- Headless modes: JSON and RPC output for automation and integrations
- Provider choice: /login selects a subscription or API-key provider on first launch
- Session management CLI: agents, attach, --resume, status, doctor, update, and shutdown commands
