Prime Intellect has launched Prime Agent, a self-improving coding harness built around two core abstractions: the Recursive Language Model (RLM) and Continual Harness. According to the company, these abstractions allow the agent to treat context as a variable and sub-agent delegation as function calls within a persistent REPL, enabling programmatic access to its own history, tools, and sub-agents. The Continual Harness abstraction treats the harness’s state — including prompts, skills, memory, and sub-agents — as something the agent can create, read, update, and delete from its own trajectory, supporting CRUD operations and agent-to-agent communication. Prime Agent runs a persistent IPython kernel as its primary tool, with other harness features invoked as functions within the kernel. Sub-agents are implemented as separate Prime Agent instances, each with its own model, kernel, session tree, and conversation history. The system includes a background daemon that manages live agent sessions over a local socket, allowing users to attach and detach without disrupting the agent loop. Session history is stored as append-only JSONL files on disk, with compaction triggered by context thresholds or manually via compact.run() to clean the main context while preserving full history accessibility. The REPL requires asynchronous garbage collection to manage IPython state and prevent memory buildup. Prime Agent supports parallel fan-out of sub-agents, mid-flight steering via agent_message.send, and orchestration across sessions through its daemon-mediated A2A messaging, limited to nuclear family sessions to prevent undesirable cross-talk. The harness is designed to work with current open and closed frontier models and is intended to scale with future model generations trained around it. Prime Agent is fully open-source and can be installed via a provided shell script.
Key facts
- Prime Agent uses a persistent IPython kernel as its only tool.
- Sub-agents are implemented as separate prime-agent instances with their own kernels and session trees.
- Session history is stored as append-only JSONL files on disk.
- Compaction can be triggered manually via compact.run() or by context threshold.
- Prime Agent supports agent-to-agent messaging via its background daemon.
- Multi-agent communication is limited to nuclear family sessions.
- Prime Agent is open-source and installable via curl command.