Prime Agent: When AI Learns to Build Its Own Tools — A Recursive Revolution

Forum topic · 小凯 · 2026-08-25

Summary

Prime Agent is a self-improving recursive language model (RLM) harness that lifts performance on the ARC-AGI-3 abstract reasoning benchmark from roughly 30% (baseline for frontier LLMs) to 95.5%. Instead of scaling model parameters, it equips the model with a persistent IPython REPL, a continual harness that retains histories, memories, skills, prompts, and subagent specifications, recursive subagents with direct agent-to-agent communication, an 'Agents View' for human oversight, and a resource-accounting system. The approach converts closed-book inference into open-book experimentation: the model writes, executes, and verifies code, reuses helper functions across tasks, and decomposes problems among specialized subagents. Beyond ARC-AGI-3, it matches or exceeds baselines on long-context coding, GPU kernel generation, simulator building, autonomous nanoGPT speedruns, and shows steady progress in Factorio. The article also cautions that the score may reflect brute-force enumeration rather than human-like abstraction, and notes costs in compute, latency, complexity, and security. It suggests the next AI breakthrough may come from smarter cognitive scaffolding rather than bigger models. Open code is available on GitHub.

Prime Agent: When AI Learns to Build Its Own Tools

*English structured summary of a Chinese forum post discussing the Prime Agent system and its results on ARC-AGI-3.*

The core idea

Traditional LLMs answer like students taking a closed-book exam: no scratch paper, no verification, no tools. Prime Agent instead gives the model a persistent lab — a Python environment, a memory system, and a team of subagents — turning inference into open-book experimentation.

Why LLMs fail at ARC-AGI-3 (30% baseline)

ARC (Abstraction and Reasoning Corpus, by François Chollet) tests fluid intelligence: inferring novel transformation rules from examples. It stresses compositional, variable-binding, topological, arithmetic, and geometric reasoning. Frontier LLMs score about 30% because they:

Are constrained by fixed context windows
Cannot execute code to verify hypotheses
Cannot iterate, sketch, save intermediate results, or delegate
Must rely purely on "mental simulation"

The five key components of Prime Agent

1. Persistent IPython REPL — a long-lived Python environment following the Recursive Language Model abstraction. The model defines reusable helper functions, builds hierarchical tool chains, and applies test-time compute: trying multiple methods, enumerating possibilities, and verifying results with scripts. 2. Continual Harness — retains histories, memories, skills (reusable code snippets), prompts, and subagent specifications across tasks, enabling continual learning rather than starting from zero each time. 3. Recursive subagents — specialized agents (coding, debugging, retrieval, verification) with direct agent-to-agent communication, no central dispatcher, all sharing the same persistent REPL. 4. Agents View — a human oversight window for observing, auditing, and intervening on agent behavior. 5. Resource accounting — tracks per-agent compute and API costs with budget caps.

Results

| Task | Result | |------|--------| | ARC-AGI-3 | 95.5% (vs ~30% baseline) | | Long-context coding | Matches or exceeds native/popular frameworks | | GPU kernel generation | Matches or exceeds | | Simulator building | Matches or exceeds | | Autonomous nanoGPT speedrun | Matches or exceeds | | Factorio | Steady technical progress + parallelization |

Caveats

The 95.5% may come from brute-force enumeration and verification with code, not human-like abstraction — a different kind of "intelligence" than a human solving the same puzzle.
Costs: higher compute, higher latency, architectural complexity, and new security surfaces (persistent code execution, subagent permissions).

Takeaway

Prime Agent suggests a shift from "bigger models" to "smarter scaffolding": instead of only scaling parameters, equipping a fixed model with tools, memory, and collaboration can produce qualitative leaps. The next AI breakthrough may lie in how we design cognitive tools for AI, not just in model size.

References

Karten, S., Zhang, A. L., Thomas, K., et al. (2026). *Prime Agent: A Self-Improving RLM Harness*. arXiv preprint.
Chollet, F. (2019). *On the Measure of Intelligence*. arXiv:1911.01547.
ARC-AGI-3 Benchmark: https://arcprize.org/
Prime Agent source code: https://github.com/PrimeIntellect-ai/prime-agent