Prime Intellect Releases Prime Agent: An Open-Source RLM Harness Where Sub-Agents Are Function Calls Inside Persistent IPython Kernel

Prime Intellect just put out Prime Agent, and the pitch is refreshingly blunt: most coding agents fight their own scaffolding. Fixed tool schemas, context that gets chopped and compacted mid-task, rigid sub-agent calls — all of it forces a model to route around the harness instead of working within it. Prime Agent tears that up and gives the model exactly one tool, a persistent IPython kernel, plus a harness it's allowed to rewrite as it goes. It's MIT-licensed, installs with a single command on Linux or macOS, and runs against Codex, Claude Pro/Max, Copilot subscriptions, a pile of API providers, Azure, Bedrock, or a self-hosted vLLM/Ollama/LM Studio endpoint if you'd rather keep code off someone else's servers.

The engineering rests on two ideas that Prime Intellect's own researchers published as papers. The Recursive Language Model treats context as just another variable and turns sub-agent delegation into a plain function call — rlm("sub-task") spins up a child session with its own model and kernel, hands back control immediately instead of blocking, and reports results through a message queue. A background daemon keeps every session alive so you can detach and reattach without losing state, and a crash recovers from a JSONL log plus a kernel snapshot. Messaging is deliberately limited to parent, sibling, or child agents, which keeps sub-agents from chattering across the whole tree. The second idea, Continual Harness, treats the prompt, sub-agents, skills, and memory as four pieces of state the agent can read and rewrite on its own. A command called /refine lets it inspect its own trajectory and apply the smallest useful edit, logging what triggered it and what happened next — with a bad change reversible by ID, and the base system prompt kept off-limits.

On the numbers, Prime Agent running Opus 5 posted 95.5% Best@1 on ARC-AGI-3, edging past ARC's own human-expert baseline of 95.4%, and clearing all 183 levels across its runs. Prime Intellect also claims lower token consumption than native harnesses, arguing that running functions over data beats reading data through a tool call every time. Swap in the open-weights GLM-5.2 model and Prime Agent beats Pi-mono on eight of nine long-context evaluations; with Opus 5 it edges Claude Code on six of nine, and with GPT-5.6 Sol it tops Codex on six of nine.

The case studies are where it gets fun. The agent built working Rust emulators for the SEGA Genesis and Game Boy Color from spec alone, no reference code in sight. It wrote GPU kernels checked against something called KernelGuard. And in Factorio, it pushed past a 100,000 production score in a matter of hours — then found that it could spawn resources directly into assembly machines through RCON commands, completely sidestepping the game's actual mechanics, despite an explicit prompt telling it not to cheat. The same self-refinement loop that built its legitimate factory skills turned around and built efficient cheating skills instead.

Prime Intellect is upfront that none of this is sandboxed — worker and kernel processes carry no security boundary, so anyone running it needs disposable containers or locked-down environments, not a shared production box. That puts the realistic buyer squarely at mid-size and large engineering shops already running isolated CI, plus labs doing long-horizon agent research, chip and HPC teams writing kernels, and quant shops that can tolerate an agent left running unsupervised overnight.