Prime Agent Review: The Self-Improving RLM Coding Harness

Key highlights:

• 16K+ stars, ~8.5K of them in a single week — one of the fastest-climbing agent repos of August 2026

• MIT-licensed, macOS and Linux, built on top of the [`pi`](https://github.com/earendil-works/pi) TUI core

• Two core abstractions: the Recursive Language Model (RLM) (context as a variable, subagents as function calls) and the Continual Harness (the agent can CRUD its own prompts, memories, skills, and subagent specs)

• Benchmark headline: 95.5% RHAE Best@1 on ARC-AGI-3 with Opus 5, just past the reported 95.4% human-expert baseline — with a scorecard published for replay

• Daemon-backed sessions keep running after your terminal disconnects; you reattach later

Prime Intellect is best known for decentralized GPU compute and the PRIME-RL / verifiers training stack. Prime Agent is a different product: not a model, but a harness — the scaffolding layer between a model and your machine.

Their argument for building a new one: modern harnesses were designed around earlier model generations. Fixed tool-calling schemas and automatic context compaction force a frontier model "to work around its own scaffolding instead of leveraging it," and hand-engineered subagents, prompts, skills, and memory are set once at design time, never adapting to what the agent learns during a run.

Prime Agent's response is to make the scaffolding programmable and mutable. Two abstractions carry that:

Most harnesses give a model 15–30 tools: read_file, edit_file, bash, grep, spawn_subagent. Prime Agent gives the model one: a persistent IPython kernel. On initialization, the kernel pre-imports every skill and tool as a Python module — including rlm for recursive subagent calls.

The practical consequence is token efficiency. If an agent needs to know how many test files import a deprecated module, a conventional harness reads files into context and reasons over them; Prime Agent writes three lines of Python and reads back a number. Prime Intellect's framing: it "saves tokens by programmatically running functions over data rather than spending tokens reading data using tools."

This is the part that made the repo trend. Spawning a subagent launches a full session — its own model, its own IPython kernel, its own session tree and history:

The headline: Prime Agent with Opus 5 scores 95.5% RHAE Best@1 on ARC-AGI-3, edging past the reported human-expert baseline of 95.4%. Across three runs it landed at 95.0, 95.2, and 95.5, with 99.97% Best@3 and all 183/183 levels complete. A median scorecard is published for action replay.

That number needs context, and Prime Intellect supplies some: no model has been trained around Prime Agent, so the gain is harness design, not model capability. For comparison, ARC Prize's verified figures for the raw models sit far lower — reporting has put Opus 5 around 30% and GPT-5.6 Sol Max in single digits. That gap is exactly why the claim is contested.

The more useful result for working developers is the long-context comparison, where Prime Agent runs the open-weights GLM-5.2 against closed harnesses running their native models:

This is a condensed version. The full article includes architecture diagrams, code examples, comparison tables, and practical use cases.

👉 Read the complete analysis on andrew.ooo

Topics: ai agent, coding agent, prime agent, prime intellect, rlm