Prime Agent: Prime Intellect's RLM Coding Harness

Prime Intellect announced Prime Agent on August 5, 2026, and open-sourced it under the MIT License the following day, August 6. The GitHub repository is PrimeIntellect-ai/prime-agent; it picked up more than 2,000 stars on GitHub Trending on launch day and had reached roughly 17.4k stars as of August 20, 2026 (official blog post: https://www.primeintellect.ai/blog/prime-agent ). At the core of the design are two abstractions: the Recursive Language Model (RLM), which governs how the agent reasons, and the Continual Harness, which governs how the agent's scaffolding evolves. For context on how it stacks up against incumbent coding agents, see the complete guide to Claude Code and our Codex vs. Claude Code vs. Cursor vs. Copilot comparison.

The defining trait of RLM is that the model uses a persistent IPython kernel as its one primary tool. Most coding agents define file reads, search, and command execution as separate tool schemas that the model calls one at a time. Prime Agent's RLM takes a different approach: it treats context as a variable. Rather than loading huge inputs — long logs, an entire repository, prior conversation history — directly into the prompt, it holds them as Python variables that the model inspects, slices, and transforms with code, pulling out only what it needs. This lets it work over large data without blowing out the context window.

The second defining trait is that sub-agent delegation is treated as a function call. Sub-agents are spun up as full sessions via await rlm("task") and return messages asynchronously. That lets the parent agent fan out multiple sub-agents in parallel, and it can steer a sub-agent mid-flight — issuing new instructions before that sub-agent finishes — rather than only waiting for a final result. This is a fundamentally different approach from the conventional combination of tool schemas plus context compression (summarizing or using RAG to trim context): RLM treats both context and control flow as code. On the parallel-execution side, it's also worth comparing this to Mosaic's approach to running Claude Code in parallel.

Because of this RLM structure, Prime Agent is well suited to (1) parallel fan-out, delegating multiple subtasks to sub-agents at once; (2) mid-flight steering, sending new instructions to a sub-agent before it finishes; and (3) long-running tasks that keep state alive inside the kernel over extended runs. On officially reported benchmarks, using Opus 5, Prime Agent achieved 95.5% on ARC-AGI-3 (RHAE Best@1), above the reported human-expert baseline of 95.4%. On EmulatorBench, it successfully reproduced Sega Genesis and Game Boy Color emulators, and on MazeBench it outperformed each frontier model's native harness. These are the only benchmark figures published by Prime Intellect to date; no other benchmark results are publicly available.

There is only one officially documented installation method. It fetches a version-pinned release, verifies its SHA-256 checksum, installs the prime-agent command, and provisions the IPython runtime. As a general precaution when running any curl | sh installer, it's worth reviewing the script contents before piping it into a shell.

This is where Prime Agent is most often misunderstood when described as "self-improving." The /refine command reviews the current trajectory (the record of what the agent just did) and applies the minimal CRUD edits needed to improve outcomes to the harness. It runs in a background planning phase that considers candidate edits and a fast apply phase that applies them, and past refine history can be rolled back. Crucially, /refine never retrains the base model. "Self-improvement" here means an explicit, persistent, and reversible change to the scaffolding — auxiliary prompts, memory, skill descriptions, reusable sub-agent specs — not a change to model weights. That persistent state lives inside the IPython kernel as rlm.harness, and by default it is session-local, meaning it applies only within that session. This idea of improving the scaffolding rather than the model itself also connects to the Loop Engineering paradigm for AI agents.

As the table shows, Prime Agent's distinguishing features are its unified design of treating sub-agents as function calls, and its ability to CRUD-edit the harness itself as persistent state. On licensing, it shares the MIT model with OpenHands and SWE-agent, but its underlying approach to context management and sub-agent control — built on the RLM abstraction — is what sets it apart.

Prime Agent is better suited to long-running autonomous tasks and research-oriented, experimental use than to quick, single-shot coding help. Its reported benchmarks — ARC-AGI-3 and MazeBench — also reflect tasks that reward exploration and iteration. A few things to keep in mind before adopting it: first, because the primary interface is a Python REPL (IPython kernel), comfort with basic Python is effectively a prerequisite. Second, as with any curl | sh installer, reviewing the script before running it is good practice. Third, the bounded autonomy mode via /autonomous puts the responsibility on the user to configure what scope of actions and permissions the agent is allowed. Fourth, because harness state is session-local by default, a harness improved via /refine in one session is not automatically shared with another session or a teammate.