Prime Agent is Prime Intellect's open-source coding and research agent, built for long-running autonomous work rather than single-session chat. It is organized around two abstractions the project treats as load-bearing: the Recursive Language Model (RLM), which turns context into variables and tools into function calls inside a persistent REPL, and the Continual Harness, which keeps supplemental prompts, memories, skill descriptions, and subagent specifications as durable state the agent can revise.

The defining architectural choice is that everything is programmatic. A persistent IPython kernel is the built-in model tool, so file operations, shell commands, tool use, subagent spawning, and context management all happen as Python code rather than as separate hard-coded tool schemas. Subagents are a first-class primitive: calling rlm(...) spawns a real child agent for parallel or background work and returns its result programmatically, which lets an agent fan out and collect results without routing every step through the user.

Self-improvement is scoped deliberately narrowly. The /refine command reviews the current trajectory and applies small, evidence-backed updates to supplemental harness state — it never rewrites the immutable base system prompt, and recorded snapshots support rollback. Refinements are local to the session by default. Skills are importable Python packages rather than prose files, and a built-in skill creator converts recurring workflows into project or personal skills.

Continuity is the other half of the design. Daemon-backed sessions keep IPython state, schedules, and subagents alive when the terminal disconnects, and can be reattached later via prime-agent attach. Heartbeats, prime-agent schedule, persistent /goal objectives, automatic compaction, and a bounded /autonomous mode with turn, token, and time budgets are all aimed at evaluation and research workloads that outlive a chat window. The maintainers are explicit that the worker and kernel processes improve lifecycle isolation and recovery but are not a security sandbox — the agent executes model-generated Python with the user's own permissions.

The TUI and agent core are built on top of the pi project, which the README credits directly. Released under MIT, the repository shipped v0.7.1 on 2026-08-07 and gained roughly 2,500 stars in a single day during this scan.