AI 代理與開源工具
Prime Agent Open-Sources a Persistent RLM Harness That Keeps Agent Memory, Subagents, and Execution Environments Outside the Conversation
Prime Agent supports long-running coding agents with a persistent IPython environment, daemon sessions, and a versionable Continual Harness. Its authors report that ARC-AGI-3 RHAE Best@1 rose from 30% to 95.5%, but the score reflects the complete model-and-harness system and should not be treated as an isolated comparison of underlying model capabilities.
The Prime Intellect research team has released Prime Agent, an attempt to address two common failure points in long-horizon agents: models have limited effective context, while terminal state, subagents, and task state can disappear when a session is interrupted. Rather than adding another layer of natural-language planning, its core is a persistent IPython REPL. Prompts, file contents, and tool results can be handled as variables, and the model can programmatically call `rlm(...)` to create recursive subagents, moving large-context processing and test-time compute out of the main conversation.
A separate Continual Harness stores history, memory, skill descriptions, supplemental prompts, and subagent specifications. `/refine` can propose small, reversible updates based on the current trajectory, but it does not rewrite the immutable base system prompt. The daemon preserves Python state, schedules, heartbeats, and running agents, allowing users to reconnect after a terminal session drops. Autonomous mode also sets limits for turns, tokens, time, and quality thresholds to avoid mistaking “resource limit reached” for task completion.
The paper claims that, under the ARC-AGI-3 RHAE setting, Prime Agent increased Best@1 from 30% to 95.5%. It also compares the system with other harnesses on long-context programming, GPU kernel generation, simulator construction, nanoGPT acceleration, and Factorio tasks. The key takeaway is that the agent framework itself can make an enormous difference, but these are still primarily evaluations conducted by the authors, and differences in models, costs, and retry budgets should not be overlooked.
The code is available under the MIT License and supports macOS and Linux. Deployers should take particular care: model-generated Python code, shell commands, and skills run with the user’s permissions, and lifecycle isolation between workers and kernels is not a security sandbox. The next things to watch are independent reproductions, token and compute cost per successful run, and whether persistent memory develops contamination or capability degradation after repeated self-revisions.