← All posts
AI Agent

My Pi Research Agents Setup

With just these 3 files, you can spawn a group of agents for research and report writing.

Sep 26, 20262 min read

How to setup (super simple. 3 files)

  1. Install Pi
  2. researcher.md Put this file to ~/.pi/agent/agents/
  3. report-writer.md Put this file to ~/.pi/agent/agents/
  4. research.md Put this file to ~/.pi/agent/prompts

Why Pi

Blazing fast, lean, and token-efficient

The /research command that delegates work to other agents

A /research command has been added to the pi setup. Type /research <topic> and you get a report, output as a Markdown file plus an HTML page. The main agent doesn't do the research itself — instead, it delegates the work to other agents and then consolidates the results. Each agent can be individually configured with the model that suits it best.

In the video, I demonstrate using inclusionai/ling-3.0-flash. It's incredibly fast.

How it works

flowchart LR
    U(["/research &lt;topic&gt;"]) e0@--> P["Main agent<br/>checks date + language"]

    P e1@--> R1["researcher 1<br/>angle A"]
    P e2@--> R2["researcher 2<br/>angle B"]
    P e3@--> R3["researcher 3<br/>angle C"]
    P e4@--> R4["researcher 4<br/>angle D"]

    R1 e5@-- brief --> W["report-writer"]
    R2 e6@-- brief --> W
    R3 e7@-- brief --> W
    R4 e8@-- brief --> W

    W e9@--> MD[/"report.md"/]
    W e10@--> HTML[/"report.html"/]

    e0@{ animate: true }
    e1@{ animate: true }
    e2@{ animate: true }
    e3@{ animate: true }
    e4@{ animate: true }
    e5@{ animate: true }
    e6@{ animate: true }
    e7@{ animate: true }
    e8@{ animate: true }
    e9@{ animate: true }
    e10@{ animate: true }

What it does

Take /research GPT-5 vs Claude 5 for coding as an example. A few minutes later, two files appear in the folder:

  • 2026-09-25-1430-gpt-5-vs-claude-5-comparison.md
  • 2026-09-25-1430-gpt-5-vs-claude-5-comparison.html

The prompt

Research: $@

1. Date. Run `date`. Never guess the date.
2. Language. Match the topic's language (Chinese stays Chinese).
3. Research. Spawn one or more `researcher` subagents, each with a
   distinct angle. Don't research yourself or answer from memory.
   Start each task with today's date, the language, and the time
   window that counts as current.
4. Report. Spawn a `report-writer` with the findings. It writes a
   timestamped .md and a matching .html.

Don't ask me to clarify scope. Make a reasonable call and proceed.

(This is the condensed version. The full file is agent/prompts/research.md.)