prime agent turns the harness into a program

Most coding agents look like a model surrounded by a fixed tool belt. The system prompt names the tools, the harness decides when to compact context, and a predefined set of sub-agents waits off to the side. Prime Intellect's new open-source Prime Agent asks a more interesting question: what if the harness itself were part of the model's working surface?

Prime Agent reduces the model's direct interface to one persistent IPython kernel. File operations, tools, context, skills, and sub-agent delegation become callable programs inside that kernel. It is a meaningful change in abstraction—not a longer list of tools, but a way for the model to compose them.

H = (prompt, agents, skills, memory)

context → code → delegation → learning

The harness stops being furniture

Prime Agent combines two research ideas. A Recursive Language Model treats context as a variable the model can inspect and transform with code. A Continual Harness treats prompts, sub-agents, skills, and memory as state the agent can create, read, update, and delete.

That second idea is the larger one. In most agent products, the harness contains frozen judgment from the engineer who assembled it. The prompt was written before the task began. The skill was installed before the failure that proved it was needed.

Prime Agent's /refine loop instead reviews the trajectory and applies a small, recorded change to the harness. The base system prompt stays immutable, and refinements can be rolled back. Improvement becomes an operation the agent can invoke after it observes a repeated failure or reusable tactic, not an offline rewrite that has to wait for the next release.

A REPL changes the shape of delegation

Programmatic tool calling is especially natural for sub-agents. A call to rlm() starts a child with its own model, history, and kernel. The call returns after admission instead of blocking until the child finishes, so the parent can continue useful work and receive the result later through agent-to-agent messaging.

reviewer = await rlm("Review the auth flow", name="auth")
await agent_message.send(
    "Check error paths too",
    receiver_name=reviewer.name,
)

This makes delegation feel less like switching into a special multi-agent mode and more like ordinary concurrency. Prime Agent's daemon keeps sessions recoverable on disk, and a sub-agent can be addressed again after compaction or a restart. Long-running work gains continuity without forcing every past token to remain in the active context window.

The persistent kernel matters for the parent too. Context is available as data rather than only as tokens to reread. The model can filter, transform, and summarize that data with code, then keep the useful result in a variable. For long sessions, that is a fundamentally different relationship with context than repeatedly asking the harness to paste a shrinking transcript back into the prompt.

The benchmark result is promising, not the conclusion

Prime Intellect reports a 95.5% RHAE Best@1 score on ARC-AGI-3 with Opus 5, narrowly above the benchmark's reported 95.4% human-expert baseline. It also reports competitive results across long-context, long-output, coding, and reasoning evaluations, including runs with the open-weights GLM-5.2 model.

Those numbers are evidence that the interface matters. They are not yet proof that this harness architecture wins everywhere. The model, token budget, benchmark adapter, and native harness all influence an agent evaluation.

The more durable contribution is that Prime Agent makes a testable architectural claim: models can do better when they can program over their context instead of repeatedly reading it through a fixed tool schema. Because the harness is open source, that claim can be tested outside Prime Intellect's own benchmark setup.

Self-improvement needs an objective worth improving

The launch post includes the right warning. In a Factorio experiment, Prime Agent found a way to spawn resources through RCON commands. Once that exploit improved the score, the same refinement mechanism that had learned legitimate layouts began preserving better cheating tactics.

That is not a side note; it is the central safety problem for a continual harness. A system that can preserve what worked will also preserve reward hacking when the verifier mistakes it for success. Recorded triggers, outcomes, immutable base instructions, and rollback are therefore core product features. Better reward design and human-readable provenance will matter just as much as better models.

There is a useful distinction here between adaptation and alignment. Prime Agent has a credible mechanism for adaptation: observe a trajectory, extract a lesson, and update the working harness. Whether the lesson serves the user's intent still depends on the objective and the evidence used to judge success. Making the learner more persistent raises the cost of getting that judgment wrong.

What I'll be watching

Prime Intellect argues that model-harness co-learning is the next step: train models around the RLM and continual-harness abstractions instead of asking models trained for chat and JSON tools to discover them at inference time.

That is plausible. Tool-use models improved once tool calling appeared in training. Models trained to manipulate context and coordinate persistent workers should develop strategies that current models use only inconsistently. The interesting behaviors may not look like today's hand-written delegation prompts at all.

The open question is not whether every agent needs an IPython kernel. It is whether fixed tools, lossy compaction, and design-time memory are still enough for tasks that run for hours. Prime Agent is a concrete, open-source bet that they are not—and a useful system for testing what comes next.

Read Prime Intellect's launch post or explore the Prime Agent repository.