Prime Agent vs Claude Code: memory and harness tradeoffs

The speaker compares Prime Agent paired with Claude Opus 5 against Claude Code, focusing on control over the coding assistant's harness. They describe Prime Agent as an open source, MIT-licensed terminal tool that reads repositories, edits files, runs commands and checks results.

Persistent sessions and learned instructions are central to the comparison. According to the walkthrough, sessions survive terminal detachment, while the refine command saves reviewable lessons as memories, supplemental prompts or skills. These changes can be rolled back; the base system prompt remains immutable. Autonomous work has configurable turn, token and time budgets.

The speaker describes connecting Opus 5 through an Anthropic API key and config file. The speaker describes a one-million-token context window and low, medium or high effort settings. They also name vLLM, Ollama and LM Studio as self-hosted backends for those who want to run models locally. That is a separate option from the Anthropic API configuration demonstrated here.

The reported 95.5% ARC-AGI-3 score needs context: the speaker says it is Prime Intellect's own result, has not been independently reproduced and does not appear on the official leaderboard. Practical drawbacks include setup work, reported code bloat and potentially expensive refinement runs. Because the agent executes Python and spawns subprocesses, the speaker cites Prime Intellect's recommendation to use a container for sensitive work.