Earth's Magnetic Field Research: MiniMax-M2.1 Subagent Orchestration

Executive Summary

This repository documents a comprehensive multi-agent research project investigating Earth's magnetic field changes and the scientific consensus on pole reversals. Using MiniMax-M2.1, we orchestrated 8 sub-agents across 3 phases to produce a definitive 9.9KB research synthesis answering the core question:

Is Earth's magnetic field deteriorating, and is there evidence for significant events (like pole reversals) in our lifetime?

Key Findings:

✅ YES: The field IS measurably changing (weakened ~10-15% since 1800s, pole accelerating toward Siberia at 55 km/year, South Atlantic Anomaly growing)
❌ NO: No credible evidence of imminent reversal—reversals are fundamentally unpredictable (Poisson process)

Architecture Overview

┌─────────────────────────────────────────────────────────────────┐
│                    ORCHESTRATION WORKSPACE                       │
│                   (Logging & Coordination)                       │
├─────────────────────────────────────────────────────────────────┤
│                                                                  │
│  ┌─────────────────┐                                            │
│  │     PHASE 1     │   6 PARALLEL SCOUTS (MiniMax-M2.1)         │
│  │   Research      │   Each with distinct research focus        │
│  │   Scouts        │                                            │
│  └────────┬────────┘                                            │
│           │                                                     │
│           ├─ scout-pole-reversal/output/findings.md             │
│           ├─ scout-field-strength/output/findings.md            │
│           ├─ scout-south-atlantic-anomaly/output/findings.md    │
│           ├─ scout-recent-papers/output/findings.md             │
│           ├─ scout-prediction-models/output/findings.md         │
│           └─ scout-historical-data/output/findings.md           │
│                                                                  │
│           ▼                                                     │
│  ┌─────────────────┐                                            │
│  │     PHASE 2     │   1 AGGREGATOR (MiniMax-M2.1)              │
│  │   Synthesis     │   Synthesizes all Phase 1 findings         │
│  └────────┬────────┘                                            │
│           │                                                     │
│           ├─ output/synthesis.md (9.9 KB)                       │
│           ├─ output/key-questions.md                            │
│           └─ output/summary.txt                                 │
│                                                                  │
│           ▼                                                     │
│  ┌─────────────────┐                                            │
│  │     PHASE 3     │   1 SYNTHESIZER (MiniMax-M2.1)             │
│  │   Final Answer  │   Produces user-facing conclusion          │
│  └────────┬────────┘                                            │
│           │                                                     │
│           ├─ output/final-answer.md (9.9 KB) ◄── FINAL OUTPUT   │
│           └─ output/executive-summary.md                        │
│                                                                  │
└─────────────────────────────────────────────────────────────────┘

Methodology

Phase 1: Distributed Research (6 Parallel Scouts)

Each scout operated independently with:

Model: MiniMax-M2.1 (thinking enabled)
Max Turns: 25 per agent
Tools: browser, bash, read, write
Logging: Append-only log.md per agent (following logging protocol)

Scout Assignments:

Scout	Research Focus	Key Questions
scout-pole-reversal	Pole flip mechanics & predictions	When might reversal occur? What triggers it?
scout-field-strength	Field decay measurements	How fast is field weakening? Accelerating?
scout-south-atlantic-anomaly	SAA & regional anomalies	What's happening in SAA? Implications?
scout-recent-papers	2024-2026 publications	What's cutting-edge research saying?
scout-prediction-models	Computational forecasts	What do models forecast for coming decades?
scout-historical-data	Paleomagnetic records	What does geological history tell us?

Research Protocol:

Each scout read its plan.md for clear objectives and success criteria
Scouts maintained append-only execution logs following the logging protocol
Each scout produced output/findings.md with full citations and source URLs
All outputs captured to output/ directory

Phase 2: Synthesis (1 Aggregator)

The aggregator synthesized all Phase 1 outputs:

Read: All 6 scout findings.md files
Process: Identified consensus points and debates
Output: synthesis.md (thematic organization) and key-questions.md

Phase 3: Final Deliverable (1 Synthesizer)

The synthesizer produced the user-facing answer:

Read: Phase 2 aggregation outputs
Output: final-answer.md with direct answer to the core question

Logging Protocol

This project strictly followed the Logging Protocol (see /Users/jay/.claude/skills/logging-protocol/):

Key Principles:

Append-Only Log: Never edit previous entries. History is immutable.
Forward-Only Fixes: Backtracking = new step that fixes forward
Atomic Steps: Each step has Pre-Execution → Execution → Post-Execution
Linear Readability: Any agent can read logs top-to-bottom without graph traversal

Log Structure Used:

# STATE
## Current
step_id: STEP-XX
status: IN_PROGRESS
objective: [current objective]

## Decisions (append-only)
- STEP-01: [decision made]

## Blockers (append-only, mark resolved inline)
- STEP-03: [blocker] → RESOLVED: [how]

---

# STEP LOG (append-only)

## STEP-01
### Pre-Execution
Objective: [what to accomplish]
Beliefs: [what we believe]
Assumptions: [what we're assuming]

### Execution
- [x] Action taken
- Finding: [what discovered]

### Post-Execution
Outcome: PASS | PARTIAL | FAIL
Belief updates: [any changes]

Where Logs Are Located:

Orchestrator Logs:

orchestrator/log.md - Master orchestration execution log
orchestrator/decisions.md - Pre-flight architectural decisions

Agent Logs:

phase1/scout-*/log.md - Individual scout execution traces
phase2/aggregator/log.md - Aggregator execution trace
phase3/synthesizer/log.md - Synthesizer execution trace

Subagent Orchestration Framework

This project used the Pi Subagent Orchestration skill (see /Users/jay/.claude/skills/pi-subagent-orchestration/):

Design Principles:

Auditability over convenience: Every action traceable
Git as source of truth: All commits form canonical record
Subagents are expensive: Spawn deliberately
Cleanup is mandatory: Kill what you spawn
Simple over clever: Boring solutions that work

Resource Management:

Each agent had --max-turns 25 to prevent runaway execution
tmux sessions used for observability (or --print-last for background)
All processes tracked and cleaned up after completion

Validation Criterion:

Before spawning any agent, we ensured:

"Would a blank agent with only plan.md understand this task as self-evident?"

If not, the plan was improved until it was.

Model Configuration

Critical Note: The --model CLI flag is IGNORED by Pi. Configuration must be set in ~/.pi/agent/settings.json:

{
  "defaultProvider": "minimax",
  "defaultModel": "MiniMax-M2.1",
  "defaultThinkingLevel": "thinking"
}

This was set before any agent spawning to ensure all 8 sub-agents used:

Provider: minimax
Model: MiniMax-M2.1
Thinking: enabled

Project Timeline

2026-01-13 21:15 - Workspace initialized, git repo created
2026-01-13 21:18 - Phase 1 spawned (6 scouts, parallel)
2026-01-13 21:26 - Phase 1 complete (48.5 KB findings)
2026-01-13 21:26 - Phase 2 spawned (aggregator)
2026-01-13 21:30 - Phase 2 complete (9.9 KB synthesis)
2026-01-13 21:30 - Phase 3 spawned (synthesizer)
2026-01-13 21:35 - Phase 3 complete (9.9 KB final answer)

Total Duration: ~20 minutes
Total Research: 48.5 KB Phase 1 + 9.9 KB Phase 2 + 9.9 KB Phase 3
Sources Cited: 48+ scientific sources (NASA, ESA Swarm, peer-reviewed journals)

File Structure

Earth magnetic field research/
├── README.md                          ◄── This file
├── dashboard.html                     ◄── Visual dashboard (45 KB, Three.js)
│
├── orchestrator/
│   ├── log.md                         ◄── Master execution log
│   └── decisions.md                   ◄── Architectural decisions
│
├── phase1/                            ◄── 6 Research Scouts
│   ├── scout-pole-reversal/
│   │   ├── plan.md                    ◄── Scout objectives & steps
│   │   ├── log.md                     ◄── Execution trace
│   │   └── output/
│   │       ├── findings.md            ◄── Research findings (10.4 KB)
│   │       └── run.log
│   │
│   ├── scout-field-strength/
│   │   ├── plan.md
│   │   ├── log.md
│   │   └── output/
│   │       ├── findings.md            ◄── Research findings (5.4 KB)
│   │       └── run.log
│   │
│   ├── scout-south-atlantic-anomaly/
│   │   ├── plan.md
│   │   ├── log.md
│   │   └── output/
│   │       ├── findings.md            ◄── Research findings (5.9 KB)
│   │       └── run.log
│   │
│   ├── scout-recent-papers/
│   │   ├── plan.md
│   │   ├── log.md
│   │   └── output/
│   │       ├── findings.md            ◄── Research findings (10.6 KB)
│   │       └── run.log
│   │
│   ├── scout-prediction-models/
│   │   ├── plan.md
│   │   ├── log.md
│   │   └── output/
│   │       ├── findings.md            ◄── Research findings (7.8 KB)
│   │       └── run.log
│   │
│   └── scout-historical-data/
│       ├── plan.md
│       ├── log.md
│       └── output/
│           ├── findings.md            ◄── Research findings (8.4 KB)
│           └── run.log
│
├── phase2/                            ◄── 1 Aggregator
│   └── aggregator/
│       ├── plan.md
│       ├── log.md
│       └── output/
│           ├── synthesis.md           ◄── Thematic synthesis (9.9 KB)
│           ├── key-questions.md       ◄── Unanswered questions
│           └── summary.txt            ◄── Executive summary
│
└── phase3/                            ◄── 1 Synthesizer
    └── synthesizer/
        ├── plan.md
        ├── log.md
        └── output/
            ├── final-answer.md        ◄── FINAL DELIVERABLE (9.9 KB)
            └── executive-summary.md   ◄── Quick read version

Research Findings Summary

Direct Answer

Yes, the field IS measurably changing:

Weakened ~10-15% since the 1800s
North Magnetic Pole accelerating toward Siberia at ~55 km/year
South Atlantic Anomaly grown significantly since 2014

No, there is no credible evidence of imminent reversal:

Reversals are fundamentally unpredictable (Poisson process)
Last reversal ~780,000 years ago (much longer than average ~450,000 year interval)
Models can only predict ~5 years ahead; cannot predict reversals

Key Statistics

Metric	Value
Field weakening rate	~6.3% per century
Surface field strength	25-65 microteslas
Pole migration speed	~55 km/year (accelerating from ~10 km/year in 1900s)
SAA growth	~½ continental Europe since 2014
Reversals in 83M years	183+
Last full reversal	~780,000 years ago
Reversal duration	1,000-22,000 years
Reversal interval (average)	~450,000 years
Prediction horizon	5 years (operational models)

Scientific Consensus

Agreed Upon (all scouts):

Field is weakening (~10-15% since 1800s)
North pole accelerating toward Siberia
SAA is real, growing, deepening
Reversals are statistically random—no prediction possible
Last full reversal ~780,000 years ago
Reversal duration: 1,000-22,000 years
Field never disappears completely
WMM/IGRF reliable ~5 years ahead

Debated Among Scientists:

Current change significance (normal vs. unusual precursor)
Laschamp event causal links
Reversal probability after long Brunhes period
Duration estimates for past reversals

Key Uncertainties

Is current behavior "normal" or "unusual"? — Open scientific question
When will the next reversal occur? — Fundamentally unpredictable
How fast will future reversals be? — Estimates vary by orders of magnitude
What will be the effects? — Some impacts known, others uncertain

Source References

Research synthesized from 48+ scientific sources including:

NASA Geomagnetism
ESA Swarm Mission Data
NOAA National Geophysical Data Center
World Magnetic Model (WMM2025)
International Geomagnetic Reference Field (IGRF-14)
Peer-reviewed journals: Nature, Science, JGR, Earth and Planetary Science Letters
Recent publications (2024-2026)

Technical Stack

Orchestration Framework: Claude Code + Pi Subagent Orchestration skill
Language Model: MiniMax-M2.1
Logging Protocol: Custom append-only execution logging
Dashboard: Three.js for 3D Earth visualization
Research Duration: ~20 minutes for 8 agents
Total Output: ~70 KB of research documents

How to Reproduce

Configure MiniMax-M2.1 in ~/.pi/agent/settings.json
Create workspace directory
Copy plan.md files for each agent
Spawn agents in parallel (Phase 1), then sequential (Phase 2, 3)
Monitor via tmux sessions or run logs
Collect outputs from each phase

Key Learnings

MiniMax-M2.1 produces high-quality research with reasoning traces
Parallel scouting dramatically reduces total execution time
Append-only logging enables perfect audit trails and agent handoffs
Phase-based architecture ensures clear separation of concerns
Explicit logging protocol prevents confusion during multi-agent coordination

License

This research is provided for educational and scientific purposes.

Repository Info

Repository: Earth-Magnetic-Field-Research-Minimax-w-subagents-in-pi-
Created: January 13, 2026
Model Used: MiniMax-M2.1
Total Agents: 8 sub-agents across 3 phases
Final Output: phase3/synthesizer/output/final-answer.md

Generated through multi-agent orchestration using MiniMax-M2.1 and the Pi subagent framework.

🤖 AI-Generated Repository

This entire repository was conceptualized, architected, and implemented by MiniMax-M2.1 - the reasoning model explicitly specified for this project.

AI Identity & Configuration

Model: MiniMax-M2.1
Thinking Level: Enabled
Provider: Minimax (official API)
Framework: Claude Code + Pi Subagent Orchestration

The model configuration was detected and inferred from the official Minimax provider API settings:

{
  "defaultProvider": "minimax",
  "defaultModel": "MiniMax-M2.1",
  "defaultThinkingLevel": "thinking"
}

What MiniMax-M2.1 Did:

Read and understood the Pi Subagent Orchestration skill documentation
Read and understood the Logging Protocol skill documentation
Designed the architecture - 3-phase pipeline with 8 sub-agents
Created all plans - Detailed plan.md files for each agent
Configured the model - Set up MiniMax-M2.1 with thinking enabled (inferred from provider API)
Spawned all agents - Orchestrated parallel and sequential execution
Monitored progress - Tracked all 8 agents through completion
Compiled findings - Synthesized 48.5 KB of research into final answer
Built the dashboard - Created visual Three.js presentation
Documented everything - Wrote comprehensive README with methodology
Pushed to GitHub - Set up repository with description and topics

Key Decisions Made by MiniMax-M2.1:

Model Selection: Used MiniMax-M2.1 as detected from official provider API
Architecture: Designed 3-phase pipeline (parallel scouts → aggregator → synthesizer)
Logging Strategy: Used append-only logging protocol for perfect auditability
Phase Distribution: 6 parallel scouts for breadth, 1 aggregator for synthesis, 1 synthesizer for final output
Research Focus: Identified 6 distinct research angles to cover comprehensively
Dashboard Design: Created visual presentation with 3D Earth animation

Transparency:

All execution logs are preserved in this repository, showing:

Pre-execution assumptions and objectives
Real-time execution progress
Findings and discoveries
Post-execution outcomes
Belief updates and hypotheses

This demonstrates complete transparency into MiniMax-M2.1's decision-making process with thinking enabled.

Repository created by MiniMax-M2.1 via Pi harness - January 13, 2026

Name		Name	Last commit message	Last commit date
Latest commit History 7 Commits
orchestrator		orchestrator
phase1		phase1
phase2/aggregator		phase2/aggregator
phase3/synthesizer		phase3/synthesizer
README.md		README.md
dashboard.html		dashboard.html

EmZod/Earth-Magnetic-Field-Research-Minimax-w-subagents-in-pi-

Folders and files

Latest commit

History

Repository files navigation