Agent Fungibility — Scalable Multi-Agent Architecture

Core Insight: "Fungibility bestows a LOT of really good properties automatically in a computer system."

The goal: Build the machine once, set it running, move on. The swarm handles the rest.

The Philosophy

All agents are identical generalists. Any agent can work on any bead. There are no "frontend agents" or "testing agents" — just agents executing beads from a shared pool.

Why this wins:

• Failure resilience — Dead agent? Start another. No role replacement needed
• Linear scaling — 3→10→30 agents requires no rebalancing
• No bottlenecks — No "critical specialist" blocking the pipeline
• Reduced human involvement — Front-load planning, then step away

---

THE EXACT PROMPT — Agent Initialization

Give this to every agent regardless of type (Claude Code, Codex, Gemini):

First read ALL of the AGENTS.md file and README.md file super carefully and understand ALL of both! Then use your code investigation agent mode to fully understand the code, technical architecture and purpose of the project. Then register with MCP Agent Mail and introduce yourself to the other agents.

Be sure to check your agent mail and promptly respond if needed to any messages; then proceed meticulously with your next assigned beads, working on the tasks systematically and tracking your progress via beads and agent mail messages.

Don't get stuck in "communication purgatory" where nothing is getting done; be proactive about starting tasks that need to be done, but inform your fellow agents via messages when you do so and mark beads appropriately.

When you're not sure what to do next, use bv to prioritize the best beads to work on next; pick the next one that you can usefully work on and get started. Make sure to acknowledge all communication requests from other agents. Use ultrathink.

Why This Prompt Works

Element	Purpose
Read AGENTS.md first	Full project context before any action
Register with Agent Mail	Coordination layer established
Check mail, respond	Prevents coordination gaps
"Communication purgatory"	Explicit anti-pattern to avoid
Use bv for next bead	Decentralized task selection
Mark beads appropriately	Status visible to all agents

---

Quick Reference

Scenario	Action
Spawn swarm	ntm spawn PROJECT --cc=3 --cod=2 --gmi=1
Initialize all	ntm send PROJECT --all "$(cat init_prompt.txt)"
Agent dies	Start new session, give same init prompt
Need to scale	Just add more: ntm add PROJECT --cc=2
Check what's ready	bv --robot-next (any agent can run this)
Claim work	br update BEAD --status in_progress

---

When to Use Fungibility

Situation	Fungible?	Why
Software development	Yes	Output matters, not discourse
Large codebase, many tasks	Yes	No specialist bottleneck
Need fault tolerance	Yes	Any agent replaces any agent
Scientific debate simulation	No	Role-based discourse IS the mechanism
Strict compliance review	No	Separation of duties required

---

Model Mix Recommendations

All agents are fungible, but models have different strengths:

Mix	Best For
3 Claude (Opus) + 2 Codex + 1 Gemini	General development
5 Claude (Opus)	Complex reasoning tasks
3 Claude + 3 Codex	High-throughput coding
2 Claude + 2 Codex + 2 Gemini	Cross-validation diversity

Start small: 3-6 agents. Scale up as coordination stabilizes.

# Recommended starting swarm
ntm spawn myproject --cc=3 --cod=2 --gmi=1

---

Core Properties

Robustness

Agents fail constantly:

• Crashes, closed tabs, context rot
• Rate limits, stuck loops
• "Euthanized" for fresh start

With fungibility: When an agent dies:

1. Bead remains marked in-progress
2. Any other agent can resume it
3. No special replacement logic
4. No dependency on that specific agent

Scaling

Agents	Specialized System	Fungible System
3→10	Rebalance roles	Just add 7 more
10→30	Complex orchestration	Just add 20 more
Role imbalance	Bottleneck	N/A — no roles

The Fountain Code Analogy

Like RaptorQ fountain codes:

• Turn work into a stream of fungible tasks (beads)
• Any agent catches any bead
• No "rarest chunk" bottleneck
• System resilient to partial failures

---

Integration with Flywheel Tools

Tool	Role in Fungibility
br	Shared bead pool — any agent reads/claims
bv	Decentralized prioritization — each agent asks "what's next?"
ntm	Spawn/manage identical agents
Agent Mail	Coordination without central dispatcher

Workflow

┌─────────────────────────────────────────────────────────────────┐
│  1. PLANNING (human-intensive)                                  │
│     └─► Create comprehensive markdown plan                      │
│     └─► Convert to beads with dependencies                      │
├─────────────────────────────────────────────────────────────────┤
│  2. EXECUTION (agent-intensive, human-free)                     │
│     └─► Spawn N fungible agents                                 │
│     └─► Each agent: read AGENTS.md → bv next → claim → work    │
│     └─► Agent Mail for coordination                             │
├─────────────────────────────────────────────────────────────────┤
│  3. MONITORING (periodic check-ins)                             │
│     └─► ntm dashboard PROJECT                                   │
│     └─► Replace dead agents as needed                           │
└─────────────────────────────────────────────────────────────────┘

---

Failure Recovery

Agent Dies Mid-Task

# Check what was in progress
br list --status in_progress

# Spawn replacement
ntm add PROJECT --cc=1

# Give same init prompt — it will:
# 1. Read AGENTS.md
# 2. Run bv, see the stuck bead
# 3. Either resume or mark blocked

Mass Failure (Context Compaction)

# All agents hit context limits simultaneously
# No panic — just restart each

ntm send PROJECT --all "/compact"
# or
# Spawn fresh: ntm spawn PROJECT --cc=5
# Give init prompt, they pick up where things left off

---

Anti-Patterns

Anti-Pattern	Problem	Fix
Assigning roles	Creates specialist bottleneck	All agents do all tasks
Central dispatcher	Single point of failure	Agents use bv independently
Agent-specific context	Replacement requires recreation	All context in AGENTS.md + beads
"Only Alice can do auth"	Blocks on Alice	Any agent reads auth docs, claims auth beads

---

Comparison

Property	Specialized Agents	Fungible Agents
Failure handling	Complex: identify role, recreate	Simple: start another
Scaling	Requires role balancing	Just add more
Replacement	Need matching specialist	Any agent works
Single point of failure	Yes (each role)	No
Coordination overhead	High	Low (beads + mail)
Human involvement	Ongoing orchestration	Front-loaded in planning

---

References

Topic	Reference
Full philosophy & rationale	PHILOSOPHY.md
Prompt variations	PROMPTS.md
Failure scenarios	RECOVERY.md
Integration patterns	INTEGRATION.md

---

Validation

A healthy fungible swarm:

• All agents running same init prompt
• No agent "owns" any domain
• Dead agents replaced without role matching
• Work continues even when agents die