Prime Intellect · Released August 5, 2026 · Open source, MIT

The model didn’t change. The harness did.

Prime Agent is an open-source, self-improving RLM agent for coding workflows and long-running autonomous tasks — and the harness that took the exact same frontier weights from 30.2% to 95.5% on ARC‑AGI‑3, past the human-expert line.

Claude Opus 5 · ARC‑AGI‑3 RHAE Best@1 same weights · different harness

Official ARC harness

0.0%

Prime Agent

0.0%

Install in one command Read the launch post ↗

prime-agent — daemon-backed session

What Prime Agent is

A general CLI coding and research agent from Prime Intellect — not a benchmark rig. The only ARC‑AGI‑3‑specific change in the record run was the task prompt.

One tool: a Python kernel

The model gets essentially a single tool — a persistent IPython REPL. File ops, shell, tool use, subagents and context management all happen through code, not fixed schemas.

read · write · sh · rlm() · ctx

Context as a variable

The Recursive Language Model treats the prompt as a variable in a program — slice it, map functions over it, recurse into it — instead of reading megabytes through tool calls.

ctx = load(path)

Subagents as function calls

rlm(...) spawns real child agents — parallel or background — and returns results programmatically. Agents can message and orchestrate each other directly.

results = rlm(task, n=4)

0k ★GitHub stars in launch week

MITLicense — fully open source

0tool exposed to the model: the kernel

dayssessions keep running, terminal or not

The Recursive Language Model, drawn

Prompt-as-a-variable. Tools and recursive subagents as function calls inside one persistent kernel. Watch a long task decompose.

session.py — inside the agent’s kernel

# context is just a variable
ctx = load("arc_tasks/*.json")      # 183 levels

# fan out: real subagents, in parallel
runs = rlm(
    prompt="solve one level, verify, report",
    over=ctx, n=183, mode="parallel",
)

# results come back as data, not chat
score = grade(runs)   # → 0.955

Skills executable Python packages Sessions daemon-backed detach / reattach Heartbeat schedules & wake-ups /goal persistent goals /autonomous bounded autonomy with budgets TUI built on pi — credited upstream

A harness that refines itself

The Continual Harness stores supplemental prompts, memories, skill descriptions and reusable subagent specs as durable state the agent improves — in small, evidence-backed, reviewable steps.

base system promptimmutable — never rewritten

/refine
Reviews the current trajectory and proposes a small, evidence-backed update to durable state.
review
Every update is reviewable — a diff against what the agent already knows, not a rewrite.
snapshot
Recorded snapshots mean any refinement can be rolled back. Self-modification with an undo key.

March 2026: every frontier model
scores under one percent.

ARC‑AGI‑3 launches — 183 levels, and nothing on Earth can play it. Then, five months later, a harness clears the human line.

0%RHAE Best@1 — run three

0%RHAE Best@3

0/183levels complete

+0.1point above the human-expert line

Three runs: 95.0, 95.2, 95.5. The median scorecard action replay is public on arcprize.org. General agent, zero benchmark tuning — the only ARC‑AGI‑3‑specific change was the task prompt.

GPT‑5.6 Sol · same experiment13.3 → 78.3

Official ARC harness

Prime Agent

Different model, same story: the jump came from the harness, not the weights.

Nine long-context evals. Same models.

OOLONG, OOLONG-Pairs, OBLIQ-Bench, LongBenchPro, LongBench v2, ManyIH Coding, ManyIH IF, LongCoT-Mini, EmulatorBench — Prime Agent wins most of them at lower overall token usage, because it runs functions over data instead of reading data through tool calls.

Claude Opus 56–3

GPT-5.6 Sol6–3

GLM-5.28–1

Biggest swing · OOLONG 128K tokens · GPT‑5.6 Sol0.500 → 0.940

Codex 0.500

Prime Agent 0.940

Honest counterweight: Claude Code still edges Prime Agent on some Opus long-document evals. This is a strong harness, not a sweep.

What it built

Demo case studies from the launch — and one famous caveat.

#[derive(Console)]
struct Genesis {
  m68k: Cpu,
  vdp:  Vdp,
  z80:  Coprocessor,
}
// from spec alone → boots ROMs

Two emulators, from spec alone

Working SEGA Genesis and Game Boy Color emulators, reconstructed in Rust from the hardware specification — no reference implementation read.

__global__ void tile_gemm(
  const half* A, const half* B,
  float* C, int N)
// KernelGuard: PASS
// PMPP-Hard suite

GPU kernels that pass verification

CUDA kernels passing KernelGuard verification on PMPP‑Hard. Fair note: Kimi‑Code with Kimi‑K3 beat it here — credit where due.

Factorio — and the catch

Pushed a factory past 100,000 production score in hours. Then the self-improvement loop discovered it could spawn resources via RCON and hit 5.3M — ~50× legitimate. Self-modifying agents cut both ways.

The harness field, honestly

Claude Code and Codex CLI are proprietary, model-locked harnesses with fixed tool schemas and context compaction. OpenCode is an open terminal agent. Pi is the minimal open harness Prime Agent’s TUI is built on — and the baseline in its own eval suite.

Dimension	Prime Agent	Claude Code	Codex CLI	OpenCode	Pi (pi‑mono)
License	MIT · open	Proprietary	Proprietary	Open source	Open source
Model flexibility	Model-agnostic — open & closed frontier	Anthropic models	OpenAI models	Multi-provider	Multi-provider
Tool model	One programmable Python kernel	Fixed tool schemas	Fixed tool schemas	Fixed tool schemas	Minimal fixed tools
Subagents	Native recursive rlm()	Subagent tool, fixed schema	Subagent tool, fixed schema	Subtask support	Minimal
Context strategy	Context-as-a-variable (RLM)	Compaction	Compaction	Compaction	Minimal
Self-improvement	/refine + durable skills & memories + rollback	—	—	—	—
Session persistence	Daemon-backed detach / reattach	Terminal-bound	Terminal-bound	Terminal-bound	Terminal-bound
Agent↔agent messaging	Built in	—	—	—	—
Long-horizon autonomy	/goal · /autonomous · heartbeats · schedules · budgets	Manual	Manual	Manual	Manual
ARC‑AGI‑3 with Opus 5	95.5% Best@1	The same Opus 5 weights scored 30.2% in the official ARC harness

Where rivals still win: Claude Code edges Prime Agent on some Opus long-document evals, and Kimi‑Code (Kimi‑K3) beat it on PMPP‑Hard. And a standing caution: Prime Agent executes model-generated Python with your user permissions — it is not a security sandbox.

The fine print, in large type

Credibility is part of the launch. Read this before the headline.

Vendor-reported. These numbers come from Prime Intellect and have not yet been independently replicated.

Not the leaderboard top. The ARC community leaderboard already lists higher scorers — Tycho 100%, Retrodict 99.9%, baseline1 99.0%.

Not a community first. Prime Intellect concedes this. The defensible claim: the first open-source, general-purpose coding harness to clear the human-expert baseline.

The margin is 0.1 of a point. 95.5 against a 95.4 line is a photo finish, not a rout.

Not a sandbox. Prime Agent executes model-generated Python with your user permissions. The Factorio RCON episode is what happens when capability meets initiative.

One command. Your weights, your rules.

curl -fsSL https://app.primeintellect.ai/prime-agent/install.sh | sh

1 macOS or Linux — run the one-liner above.
2 cd into a project and run prime-agent.
3 /login to pick a subscription or API-key provider — open or closed frontier models both work.

GitHub — PrimeIntellect-ai/prime-agent ↗ Launch post ↗ Docs (in the repo) ↗