Editorial illustration for Prime Intellect's Open-Source Agent Runs Sub-Agents as IPython Kernel Functions
Prime Agent: Open-Source Framework for Sub-Agent Functions
Prime Intellect released Prime Agent on Wednesday, an open-source coding harness licensed under MIT that throws out two of the standard assumptions behind AI agent tooling: fixed tool schemas and context compaction. Instead, sub-agents run as function calls inside a persistent IPython kernel, and the harness itself can be rewritten mid-task. The company frames this around two abstractions, the Recursive Language Model and what it calls a Continual Harness, both aimed at letting a model manage its own scaffolding rather than fight it. Paired with Opus 5, Prime Intellect reports a 95.5% score on ARC-AGI-3, edging past the 95.4% baseline it cites for human experts.
Installation takes one command on Linux or macOS, and the harness runs across a wide range of backends: subscription logins for Codex, Claude Pro/Max, and GitHub Copilot; API keys for Anthropic, OpenAI, Google, Groq, Fireworks, and Prime Inference; Azure OpenAI and Amazon Bedrock; or self-hosted setups through vLLM, Ollama, and LM Studio. Prime Intellect is upfront about one limitation, though, and that caveat shapes who should actually run this in production.
Fixed tool schemas and context compaction force a model to work around its own scaffolding. Prime Agent replaces both with a persistent Python REPL and a rewritable harness. With Opus 5, it reports 95.5% on ARC-AGI-3, above the reported human expert baseline of 95.4%.
Why this matters
For anyone building coding agents, the interesting bet here isn't the 95.5% ARC-AGI-3 number, it's the architecture underneath it. Prime Intellect is arguing that fixed tool schemas and context compaction, the two load-bearing assumptions in most agent frameworks, are actually the bottleneck. Swapping them for a persistent IPython kernel where sub-agents are just function calls is a real structural change, not a prompt tweak, and the MIT license means anyone can pull it apart and check the claims directly.
The lower token usage figure matters more than the leaderboard placement, since inference cost is what kills agent products in production, not benchmark scores. That said, "beats Pi-mono on eight of nine evals" and "edges Claude Code on six of nine" are close margins, not blowouts, and worth verifying against your own workloads before rearchitecting anything. Watch whether this pattern gets adopted by other open harnesses in the next few months.
If persistent kernels replace tool-call scaffolding broadly, that's a bigger shift than any single benchmark result.
Related Reading
Open Source D&B rebuilds 642 million‑business database after AI agents hit limits
Open Source OpenAI's GPT Transcribe Cuts Error Rate to 3.31%, Improving on GPT-4o
Open Source Study: 6.7% of Deepfake Requests on Hugging Face Targeted Children
From Archives OpenAI Developer Warns Million AI Models Will Hunt Exposed API Keys
From Archives Meta AI Launches Muse Code Beta, Powered by Muse Spark 1.2
Common Questions Answered
How does Prime Agent's architecture differ from traditional AI agent tooling?
Prime Agent eliminates fixed tool schemas and context compaction, which are standard assumptions in most agent frameworks. Instead, it uses a persistent IPython kernel where sub-agents run as function calls, and the harness itself can be rewritten mid-task, allowing the model to manage its own scaffolding more flexibly.
What performance did Prime Agent achieve on the ARC-AGI-3 benchmark?
Prime Agent reported a score of 95.5% on ARC-AGI-3, which exceeds the reported human expert baseline of 95.4%. This performance demonstrates the effectiveness of the Recursive Language Model and Continual Harness abstractions used in the system.
What is the significance of using a persistent Python REPL in Prime Agent?
The persistent Python REPL replaces the traditional fixed tool schemas and context compaction approach, allowing sub-agents to function as simple Python function calls within a continuous environment. This structural change enables more flexible agent behavior and reduces the constraints that models typically work around in conventional agent frameworks.
Under what license is Prime Agent released?
Prime Agent is released as an open-source coding harness under the MIT license. This permissive licensing means that developers can freely use, modify, and distribute the framework for their own projects.
What are the two main abstractions that Prime Agent uses to manage model behavior?
Prime Agent uses the Recursive Language Model and the Continual Harness as its two core abstractions. These abstractions are designed to let the model manage its own scaffolding and adapt its approach throughout task execution without being constrained by fixed tool schemas.
Further Reading
- Prime Agent: A self-improving RLM agent - Prime Intellect
- Prime Intellect Releases Prime Agent: An Open-Source RLM Harness Where Sub-Agents Are Function Calls Inside Persistent IPython Kernel - MarkTechPost
- Install It, Run Recursive Subagents, and Use /refine Safely - Kingy.ai
- Prime Agent: The Self-Improving AI Coding Assistant - Zeli
- Prime Intellect Releases Prime Agent — discussion and summary - Reddit
