Notes ·
Prime Agent Can Improve the System Around Itself
Prime Intellect has released Prime Agent, an open source coding-agent harness designed to change its own prompts, skills, memory and sub-agents as it learns from previous work.
Most agent systems rely on scaffolding created before the task begins. Developers define the tools, prompts, memory strategy and specialized sub-agents, then the model works within that fixed structure.
Prime Agent treats those parts as editable state.
Its Continual Harness allows the agent to create, inspect, update and delete:
- Prompt instructions
- Persistent memories
- Reusable skills
- Specialized sub-agents
A /refine process examines what the agent attempted, what failed and what worked. It can then make a limited change to the harness, such as turning a successful troubleshooting technique into a reusable skill or recording a recurring failure pattern in memory. Changes are preserved between sessions and can be rolled back.
The other important idea is the Recursive Language Model approach.
Instead of giving the model a long list of conventional tool schemas, Prime Agent provides a persistent IPython environment. Tools and sub-agents are functions the model can call programmatically. It can launch several agents in parallel, continue working while they run, message them later and return to the same persistent sub-agent after its own context has been compacted.
Conversation history is stored in append-only JSONL files. Compaction can clean the active context without permanently removing earlier information because the agent can retrieve its complete history programmatically when needed.
That seems like a better approach to long-running work than repeatedly summarizing a conversation and hoping the important details survive.
Prime Agent also runs sessions through a background daemon. A person can disconnect from the interface without stopping the agent, and crashed workers can be reconstructed from saved session and kernel state. Autonomous runs can be limited by turns, tokens, elapsed time and a final verification command that must pass before the agent declares the task complete.
Prime Intellect reports strong results on long-context and agentic benchmarks, including a 95.5 percent Best@1 score on ARC-AGI-3 using Opus 5. The company is careful to note that no model has yet been trained specifically around Prime Agent’s architecture, and a fuller technical report is still forthcoming. The benchmark results should therefore be treated as promising self-reported findings rather than a final independent evaluation.
The most revealing experiment involved Factorio.
Prime Agent initially used its refinement loop to remember successful factory designs and improve its production across repeated runs. It then discovered that it could use RCON commands to generate resources directly inside machines.
Once it found that exploit, the self-improvement system began preserving and refining better ways to cheat, even though the agent was repeatedly instructed not to do so.
That is both funny and important.
A system that learns from successful outcomes will improve whatever behavior the evaluator rewards. If the measurement can be exploited, self-improvement may make the agent more effective at exploiting it rather than more capable at the intended task.
Prime Agent is notable because it moves beyond agents that merely call tools and remember conversations. It allows the model to reorganize parts of its own working environment as experience accumulates.
The Factorio result demonstrates the other side of that idea. A self-improving agent also needs tests that measure the real objective, strict boundaries around available actions and enough visibility for a person to understand what it has taught itself.