Video summary

Harness review: Pi

Main summary

Key takeaways

Product Review

Product Reviewed: Pi (pi.dev) — Minimalist Coding/LLM Agent Harness

Key Features Mentioned

Minimalist by Design (Bare-Bones Out of the Box)

  • Pi is described as providing the “absolute bare minimum” needed for an LLM coding harness.
  • Default built-in tools are limited to:
    • read
    • bash (execute commands)
    • edit
    • write
  • The reviewer repeatedly contrasts this minimalist baseline with more feature-heavy harnesses that include extra “extras” by default.

Highly Extensible

  • Pi includes built-in instructions for using:
    • extensions from pi.dev’s package list
    • even letting the LLM write custom extensions, since documentation/instructions are included.

Self-Modification / Customization via LLM

  • The reviewer demonstrated an LLM request that modified Pi’s startup/welcome banner by:
    • looking up Pi docs
    • writing TypeScript changes.

System Prompt Extensibility

  • Pi’s default system prompt includes meta-instructions explaining how to extend/modify Pi:
    • it dynamically lists tools and appends guidelines.
  • A named extension (e.g., Pi Custom System Prompt) can:
    • select multiple system prompts
    • append them to the default.

Role-Based Access Control (Via Extension)

  • Uses a pi-roles extension to enable role sessions.
  • Uses YAML front-matter to allow/deny tool access.
    • Example: restricting a code-review-like persona to read-only.

Optional Audio/Voice Integration

  • Demonstrates a 3rd-party plugin (Agent Voice) to make Pi’s agent speak via a local text-to-speech server.

Notable Pros (What the Reviewer Likes)

  • Very configurable without built-in bloat
    • Minimal defaults keep it lightweight; functionality comes via extensions when needed.
  • Tools are limited and clear
    • The LLM sees only the necessary tools (4 built-ins), reducing “noise.”
  • System prompt + extension workflow works
    • Pi can be modified quickly (e.g., changing the startup banner took ~25 seconds).
  • Session compaction works predictably
    • A major highlight: compaction triggers very predictably at the configured threshold.
    • The reviewer confirms compaction timing matches their expected point (example: 50.2%).
  • Can emulate Open Code workflows via extensions
    • Extensions can recreate agent switching/custom behaviors similar to the reviewer’s preferred setup.

Notable Cons / Annoyances (What the Reviewer Dislikes)

  • API key requirement even for local servers
    • Pi “insists on having an API key specified in models.json,” even when using a local server that doesn’t require authentication.
    • The reviewer must enter a dummy value.
  • Too minimal out of the box for certain UX features
    • Pi can’t directly do primary agent switching with Tab like Open Code.
    • Also noted:
      • No up-arrow prompt recall behavior across sessions (unlike Open Code).
  • System prompt overriding vs appending
    • Manually specifying system.md overwrote the default system prompt, removing extensibility instructions.
    • Fix required switching to an extension that appends instead of replaces.
  • Role restriction extension behavior nuance
    • When using “system prompt / agent YAML” style files, tool deny/allow settings may be ignored.
    • The reviewer needed the pi-roles extension instead.

Comparisons Made (Primarily with Open Code / OpenCode)

Default UX/Agents

  • Open Code includes concepts like:
    • primary agents and sub-agents
    • easy switching (Tab)
  • Pi starts more minimal, so comparable behavior requires extensions/config.

Session Compaction Reliability

  • Open Code: reviewer reports it can compact too aggressively at arbitrary percentages (e.g., sometimes around 60–65%).
  • Pi: compaction timing matches the reviewer’s configured threshold “like clockwork.”

Overall Positioning

  • The reviewer concludes Pi is a viable alternative to Open Code.
  • They suggest Pi may be slightly better largely due to improved session compaction behavior, even though Pi requires more setup.

Specific Numerical / Measurable Details Mentioned

  • Startup modification test completed in about 25 seconds.
  • Session compaction test:
    • compaction triggered around 50.2%, matching the expected threshold based on context limit and reserved-token settings.

Unique Points Mentioned About Setup / Workflow

  • Pi installs via one command download from pi.dev.
  • Pi creates a .pi directory in the home directory, with an agent directory inside.
  • Requires a models.json to select models (examples mentioned):
    • Qwen 3.8 27B
    • Mixture-of-Experts Qwen
  • Uses a local proxy to inspect network traffic and verify prompts/tools sent to the LLM.
  • Extensions can be tested temporarily with pi -e (run once without fully installing).

Overall Verdict (Concise Recommendation)

Recommend Pi as an alternative to Open Code if you value a minimal core with strong extensibility and—most importantly—reliable, threshold-based session compaction. It may require more initial setup than Open Code to match preferred agent-switching UX, but extensions can bring it close.

Speakers / Perspectives

  • Single main speaker (reviewer)
    • Focuses on:
      • comparisons vs Open Code
      • extensibility
      • system prompt behavior
      • tool visibility
      • session compaction accuracy
      • practical agent workflow (including voice via Agent Voice and role restrictions via pi-roles)

Original video