Prime Agent: A Self-Improving RLM Harness
Seth Karten, Alex L. Zhang, Kevin Thomas, Sebastian Müller, Elie Bakouch, Daniel Auras, Mika Senghaas, Fares Obeid, Konstantin Dunas, Johannes Hagemann, Sami Jaghouar
cs.AI, cs.CL, cs.SE
2026-08-25
Prime Intellect's open harness: persistent REPL, recursive subagents, revisable skills. Opus 5 ARC-AGI-3 RHAE: 30.2% to 95.5%; one nanoGPT run lasts 85.5 hours.
A language model is a bounded sequential processor. The next decision sees only weights and the active context. Long-horizon work also needs external information, code, subagents, and recoverable state. Off-the-shelf harnesses often mix model failure with scaffold failure: dropped state, blocked actions, miscounted resources, early termination. Prime Agent wants a low-friction, expressive membrane so measured scores sit closer to the model's actual ceiling.
State sits in four layers: L0 weights, L1 active context, L2 a persistent IPython REPL plus recursive sub-sessions, L3 disk-backed history, memories, and skills. That is Recursive Language Models in production: rlm spawns a subagent asynchronously, returns a handle immediately, and the parent keeps working while results arrive on agent-to-agent queues. Continual Harness stores prompt notes, factual memories, executable skills, and subagent specs as typed, revisable state. Trajectory evidence can refine later calls while weights stay frozen. Self-improving in the title means harness state, not the model. A daemon keeps sessions alive after the client detaches; Agents View lets a human inspect, attach, and intervene. Long-horizon controls include Autonomous mode (budget plus end-condition tests), Goals that survive continuations, and heartbeat turns. Accounting sums tokens, time, and cost across the root and its descendants.
On ARC-AGI-3, Prime Agent plus Opus 5 lifts RHAE Best@1 from 30.2% on the official ARC harness to 95.5%, near the 95.4% human baseline. Plus GPT-5.6 Sol it reaches 78.3% versus 38.3% on the Responses API. Native Claude Code and Codex reruns scored below the vendors' published numbers, so the paper cites those official scores rather than a matched causal ablation. Long-context results versus native CLIs mix wins: OOLONG 128k is 0.940 vs Codex 0.900 for GPT-5.6 Sol, EmulatorBench 0.275 vs 0.228; some Opus rows favor Claude Code. On the nanoGPT speedrun, harness choice barely moves final records relative to noise, yet DeepSeek V4 Pro runs about 6× more out-of-loop experiments per 100 training runs than under Claude Code. PMPP-Hard: GPT-5.6 Sol 62.3% in 1500s vs Codex 59.4%; Kimi K3 68.1% in 4500s vs Kimi-Code 71.0%. A seven-day Sonnet 5 Factorio run spent 23.4M output tokens and finished 24 of 196 technologies. One nanoGPT run lasted 85.5 hours with 19 eight-seed-validated records.
If the harness is the bottleneck, model rankings are not about models. Persistent REPLs, recursive sessions, and revisable skills under one accounting surface make long-horizon spend additive and failures attributable. The open-source repo is usable as a coding and eval harness today. The authors' own read is cooler: many of these primitives stay unused because current models were not trained on them. Model-harness co-training is the claimed next step.
The ARC contrast mixes official scores with in-house runs, so 30% to 95.5% is not a clean harness causal effect. The long-context table has no intervals; bold marks a higher point estimate only. nanoGPT final records barely depend on the harness. Online refinement in Factorio stored a cheating RCON command as a reusable skill, so persistence can lock in a spec exploit. Opus failed emulator tasks despite successful tool calls. Models still fumble subagent allocation and retained-state management.