Pi vs DeepSeek Harness: two open agent frameworks with opposite design philosophies
solyarisoftware · x · 2026-09-01
A Chinese blogger compares two MIT-licensed open-source agent frameworks, Pi and DeepSeek Harness (dsh):
- Pi guards the loop: read/write/edit/bash built in with a very short system prompt; Skills, Extensions and Packages stay external; Loop and Session remain accessible — like a bare shell with clear load-bearing walls.
- dsh dissolves the loop: model, tools, sandbox, storage, UI and even the agent's next reasoning step are plugins; "Everything is a Plugin" is architecture, not marketing — like a Lego studio where everything is swappable.
- Complexity placement differs: Pi leaves choice to the user so context stays clean; dsh stacks Profiles where later layers override earlier ones — fast to swap capabilities but harder to debug.
- Benchmark anecdote: running the same DeepSeek V4 Pro across Claude Code, dsh, Hermes, Pi and OpenCode on agent tasks, Pi completed the most tasks while dsh was the cheapest.
A quoted thread also covers Plannotator, a Pi plugin adding visual plan/code review in the browser (annotate plans, review diffs like a PR, distill comments into reusable Review Skills), at the cost of a heavier workflow.
More from coding & agent
- Introducing Loupe: An open-source PR review agent — andersonbcdefg · 2026-09-01
- Why "it feels better" isn't good enough for production LLM decisions — camerongreen95 · 2026-09-01
- Hermes Agent monitors chats and emails to auto-create tasks and draft work — intellectronica · 2026-09-01
- Free Australian business-day MCP server released — Impossible_Bit_2676 · 2026-09-01
- If starting with AI agents today, what would you automate first? — omnidimension85 · 2026-09-01
- SparkLLM open-sources X2.5 4B/1.7B on-device models with native 1M-token context — vllm_project · 2026-09-01