Source: MarkTechPost — 2026-08-06
Summary
Prime Intellect open-sourced Prime Agent, an MIT-licensed coding and research harness that throws out the usual tool-schema-plus-context-compaction design in favor of giving the model a single persistent IPython kernel as its only interface to the world. Tools, skills, and sub-agents all become ordinary Python function calls inside that kernel, built on two paired ideas the team calls the Recursive Language Model (context as a variable the model can program over) and the Continual Harness (a self-modifiable state the agent can create, read, update, and delete). Paired with Opus 5, Prime Agent scored 95.5% on ARC-AGI-3, edging past the reported 95.4% human-expert baseline, with Prime Intellect also reporting gains across other models versus their proprietary harnesses.
Key Takeaways
- Instead of a fixed JSON tool schema and a context-window-management layer bolted on top, Prime Agent's only tool is a persistent IPython kernel — the model writes and executes Python to call tools, spawn sub-agents, and manipulate its own conversation history as data.
- The Recursive Language Model (RLM) piece treats context itself as a variable in that kernel, so the agent can slice, summarize, store, or discard pieces of its own history programmatically rather than relying on an external compaction routine.
- The Continual Harness piece makes the harness's own configuration — prompts, skills, memory, sub-agent definitions — something the agent can CRUD from inside its own trajectory, which is the mechanism behind the "self-improving" label.
- Sub-agent delegation and agent-to-agent messaging both run through the same function-call interface: any Prime Agent session can message another persistent sub-agent the same way it would call a tool.
- On ARC-AGI-3, Prime Agent plus Opus 5 hit 95.5%, above the reported 95.4% human-expert baseline, and Prime Intellect says the improvement generalizes across models rather than being tuned to one benchmark.
- The full harness is released under the MIT license and runs on top of whichever open-weight or proprietary model the user supplies.
Reel Script
Hook: Most AI coding agents manage their memory by quietly deleting parts of the conversation and calling every tool through a rigid schema. One new open-source harness just deleted both of those ideas — and scored higher than human experts on a benchmark built to resist exactly this kind of agent.
Core Concept: Conventional agent harnesses juggle two separate systems: a tool-calling layer that describes every available function in a strict schema, and a context-compaction layer that quietly summarizes or trims the conversation once it gets too long for the model's context window. Prime Agent, released open-source by Prime Intellect, collapses both into one thing — a persistent IPython kernel that's the model's only tool. Instead of calling a tool through a schema, the model writes Python that calls it directly. Instead of an external system deciding what to compact, the model's own conversation history is just a variable it can read, slice, and store inside that same kernel — an idea Prime Intellect calls the Recursive Language Model, because the model is effectively writing and running a program over its own memory. Layered on top is what they call the Continual Harness: the agent's prompts, skills, and sub-agent definitions are themselves data the agent can edit from inside its own execution, which is what makes the system capable of modifying its own behavior over a long-running task instead of staying frozen in its starting configuration.
Hands-On: The concrete artifact worth putting on screen is the architecture swap itself, drawn as two stacks side by side. Conventional stack: model to JSON tool schema to tool executor, plus a separate context-compaction service watching the token count and summarizing history behind the model's back. Prime Agent stack: model to one persistent IPython kernel, full stop — tool calls are Python function calls, sub-agents are Python function calls that spawn and message other kernel sessions, and "compaction" is just the model assigning a chunk of its own history to a variable or writing it to disk. The other artifact worth showing is the benchmark result: Prime Agent paired with Opus 5 scored 95.5% on ARC-AGI-3, a benchmark specifically designed to test novel reasoning rather than memorized patterns, against a reported human-expert baseline of 95.4%. Prime Intellect frames the score as evidence the harness itself — not just the underlying model — is doing real work, since they report similar gains when swapping in other models against those models' own proprietary harnesses.
Takeaway: Collapsing tool-calling and context management into one programmable kernel is a genuinely different bet than the schema-plus-compaction stack most agent frameworks ship today, and an MIT license means anyone can go stress-test whether it holds up outside a benchmark built for it. Worth cloning the repo and running your own long-horizon task through it before assuming your current harness architecture is the ceiling.