Claude Code and Codex have been the mainstay agent harnesses, with OpenCode and Pi as the popular open-source alternatives. But two new entries are pushing what a harness even is: DeepSeek’s Cordis kernel, where everything — down to the agent loop itself — is a plugin you can hot-swap at runtime, and Prime Agent, an RLM-based harness that treats sub-agents and context as function calls in a live REPL. We break down what actually makes each one different.
Full shownotes at fragmentedpodcast.com.
Show Notes #
The mainstay harnesses: Claude Code, Codex, OpenCode & Pi ##
- Claude Code - Anthropic’s agentic coding tool, the default harness for many of us
- Codex - OpenAI’s coding agent
- OpenCode - the popular open-source harness with plugins
- Pi - the minimal, extensible harness; extension points let you change basically anything, including the UI
DeepSeek Harness: everything is a plugin ##
- deepseek-ai/deepseek-harness - the open-source harness; a tiny Cordis kernel that only mounts and unmounts plugins
- Cordis primer - how the plugin framework manages dependencies on an event bus
- Architecture - tool calls, models, sessions, sandboxing, scheduling, even the UI: all plugins
Prime Agent: RLMs and the continual harness ##
- Recursive Language Models - the paper behind RLMs: the context and sub-agent delegation are variables and function calls inside a live REPL
- PrimeIntellect-ai/prime-agent - the RLM-based harness; a continual environment where skills, memories, and sub-agents are CRUD operations
- Long-horizon tasks - what the continual harness unlocks: the harness improves while the work is being done
Get in touch #
Please check our contact page for the best ways to reach out to us.
We want to hear all the feedback: what’s working, what’s not, topics you’d like to hear more on.
