Multi-Agent Research Mission Control

Institutional coordination architecture for parallel AI research agents — implemented as pure context engineering in a single Markdown file.

This repository is organized around one file: CLAUDE.md. It is not configuration. It is the research artifact — a complete coordination protocol encoded as context engineering. Everything else documents, evolves, or visualizes it.

The Problem

Complex research programs exceed the context window of any single AI agent. Fully autonomous multi-agent systems fail at coordination, quality control, and research judgment. Naive parallelization produces parallel garbage: agents that can't see each other make divergent decisions, duplicate work, and miss cross-project patterns.

The missing layer is not more automation. It is institutional process design — the kind of organizational architecture that real-world mission control centers, dispatch operations, and coordination units have refined for decades.

The Approach

Multi-Agent Research Mission Control transfers institutional coordination patterns to AI agent teams. A human operator acts as the active control layer between isolated, specialized coding agents ("Lanes").

                    ┌─────────────────────┐
                    │   Mission Control    │
                    │   (Human Operator)   │
                    │                      │
                    │  · Situational       │
                    │    Awareness         │
                    │  · Dependency        │
                    │    Detection         │
                    │  · Cross-Lane        │
                    │    Synthesis         │
                    │  · Stagnation        │
                    │    Escalation        │
                    └──┬────┬────┬────┬────┘
                       │    │    │    │
              ┌────────┘    │    │    └────────┐
              ▼             ▼    ▼             ▼
         ┌─────────┐ ┌─────────┐ ┌─────────┐ ┌─────────┐
         │ Lane 1  │ │ Lane 2  │ │ Lane 3  │ │ Lane 4  │
         │ Project │ │ Project │ │ Project │ │ Infra   │
         │         │ │         │ │         │ │ (Vault) │
         └─────────┘ └─────────┘ └─────────┘ └─────────┘
              ╳            ╳          ╳            ╳
         No agent sees any other agent.
         The overall picture exists only at Mission Control.

The entire method is specified as a CLAUDE.md file — a context engineering document that gives the coordination agent its structure, protocols, and memory. No code framework. No cloud infrastructure. No agent-to-agent communication.

Why Not AutoGen / CrewAI / LangGraph?

	Agent Frameworks	Mission Control
Coordination model	Agent-to-agent (fragile, loop-prone)	Human-as-Dispatcher (robust, judgment-capable)
Complexity model	Software architecture	Institutional process design
Setup	Python environment, API keys, configuration	One Markdown file
Domain knowledge	Must be programmed	The researcher brings it
Error handling	Retry logic, guardrails	Human judgment under ambiguity
Vendor lock-in	Framework-specific	Any LLM with system prompt capability

The thesis: As agent teams take on more complex tasks, they need correspondingly complex organizational structures — and these structures come from institutional knowledge, not from software patterns.

Detailed comparison: RESEARCH-COMPARISON.md

Core Functions

#	Function	What it does
1	Status Tracking	Maintains the state of every Lane (Mode, Next Step, Blocker, Horizon)
2	Task Formulation	Translates operator intent into self-contained, copy-paste-ready agent instructions
3	Dependency Detection	Flags when one Lane produces output another Lane needs
4	Prioritization	Ranks by deadline proximity, unblocking potential, executability, value/effort
5	Synthesis	Identifies cross-Lane patterns that no single agent can see
6	Stagnation Detection	Flags missing progress, recommends action
7	Quality Control	Catches duplicates, inconsistencies, conflicting instructions

Autonomy Level

Mission Control operates at Level 4 (Approver) of the autonomy framework (Wang et al., 2025):

The agent acts autonomously; the human approves critical decisions.

This is a design decision, not a technical limitation.

Specifications

Version	Status	File	Key additions
v0.4	Released	CLAUDE.md	Operator Query Phase, Synthesis, Infra Lanes, Query Protocol
v0.5	Released	CLAUDE-v05.md	Decision Log, Answer Cache, Batch Update, Lane Lifecycle, Typed Dependencies

Repository Structure

multi-agent-research-mission-control/
│
│  ★ The Research Artifact
├── CLAUDE.md                  ← v0.4 specification (the core)
├── CLAUDE-v05.md              ← v0.5 specification (evolution)
│
│  Documentation
├── README.md
├── SYSTEM-PROMPT-DESIGN.md    ← Why CLAUDE.md is designed the way it is
├── journal.md                 ← Research journal (chronological)
├── V05-DESIGN.md              ← Design rationale v0.4 → v0.5
│
│  Academic Output
├── WHITEPAPER.md              ← Full whitepaper (German)
├── WHITEPAPER-OUTLINE.md      ← Whitepaper outline
├── RESEARCH-COMPARISON.md     ← Framework comparison
│
│  Visualization Layer (optional)
├── UI-CONCEPT.md              ← Web dashboard concept
└── forschungsleitstelle-ui/   ← Dashboard prototype
    ├── index.html
    ├── css/
    ├── js/
    └── data/
        └── demo-state.json

Quick Start

Use the method (no UI)

1. Copy CLAUDE.md as system prompt into a Claude chat
2. Describe your active projects (Lanes)
3. Mission Control tracks status, formulates instructions, detects dependencies

Run the dashboard prototype

cd forschungsleitstelle-ui
python -m http.server 8000
# or: npx serve .

Open http://localhost:8000. The dashboard loads a demo scenario.

Whitepaper

The full whitepaper (German) describes theory, architecture, and positioning:

WHITEPAPER.md — Die Forschungsleitstelle: Human-as-Dispatcher Context Engineering für parallele AI-Forschungsprojekte

License

TBD

Author

Christopher Pollin — DigitalHumanitiesCraft