SwarmOS — Darwinian AI Agent Swarm on Solana

Every death makes the swarm smarter.

What is SwarmOS?

SwarmOS is a Darwinian AI agent swarm protocol on Solana. A parent agent spawns a population of child agents to compete on a task. They are scored, evaluated, and the weakest are terminated on-chain. Before respawning successors, the system reads the failure memory of every dead agent — stored immutably in Solana account state — and injects it into the next generation's context.

The swarm evolves. Each generation is smarter than the last because it knows exactly how the previous one died.

This is evolution, not just automation.

The Problem

Multi-agent AI systems today have no memory of failure. When an agent fails, its mistakes vanish. The next agent starts from zero. At scale — thousands of agent invocations, thousands of failures — this is a compounding waste. There is no learning. There is no selection pressure. There is no evolution.

SwarmOS fixes this. Every agent death is an on-chain event. Every failure is recorded in a LineageMemory PDA. Every successor is born knowing what killed its predecessors.

How It Works

ParentAgent
    │
    ├─► ChildAgent #1 ──► ScoringOracle (pays x402) ──► score: 78 ✓ SURVIVED
    ├─► ChildAgent #2 ──► ScoringOracle (pays x402) ──► score: 45 ✗ TERMINATED
    │                                                        │
    │                                              LineageMemory PDA written
    │                                          (failure reason hash on-chain)
    │
    └─► ChildAgent #3 (Successor of #2)
            └─► Reads LineageMemory PDAs
            └─► Prompt injected with: "Agent #2 failed because: [reasons]"
            └─► Evolved behavior

The Darwinian Feedback Loop

Spawn: ParentAgent calls spawn_agent — creates an Agent PDA with generation counter and task assignment
Execute: ChildAgent runs its task using Claude AI, produces output
Pay & Score: ChildAgent pays $0.01 USDC via x402 to ScoringOracle, receives evaluation score
Evaluate: evaluate_and_prune instruction runs — agents below threshold (default 60/100) are marked Terminated, LineageMemory PDA is written with failure hash
Remember: Survivors and parent read all LineageMemory PDAs for their task type
Evolve: respawn_successor creates new agents with lineage_hash in their account, parent injects failure summaries into their context before they execute

On-Chain Architecture

Built with Anchor on Solana. All agent state lives on-chain.

Accounts

Swarm PDA — [b"swarm", authority]

authority: Pubkey
name: String
generation: u64
active_agent_count: u32
total_spawned: u64
scoring_threshold: u8     // default 60
treasury: Pubkey          // USDC token account for agent funding
task_type: TaskType

Agent PDA — [b"agent", swarm, agent_id_le_bytes]

agent_id: u64
swarm: Pubkey
parent_id: Option<u64>
generation: u64
task_type: TaskType
status: AgentStatus       // Active | Scored | Survived | Terminated | Respawned
score: u8
lineage_hash: [u8; 32]    // hash of failure context from prior generation
spawn_timestamp: i64
termination_timestamp: Option<i64>

LineageMemory PDA — [b"lineage", swarm, agent_id_le_bytes]

agent_id: u64
swarm: Pubkey
generation: u64
task_type: TaskType
failure_score: u8
failure_reason_hash: [u8; 32]   // SHA256 of failure summary (stored off-chain)
arweave_uri: String              // failure summary full text on Arweave
timestamp: i64

Instructions

Instruction	Description
`initialize_swarm`	Create Swarm PDA, set config
`spawn_agent`	Create Agent PDA, assign task, read lineage
`submit_score`	Called by oracle after x402 payment confirmed
`evaluate_and_prune`	Terminate below-threshold agents, write LineageMemory
`respawn_successor`	Spawn new Agent inheriting lineage of a terminated agent
`fund_treasury`	Accept USDC into Swarm treasury (receives LI.FI bridge output)

x402 — Agent-to-Agent Micropayments

Every ChildAgent pays before it can be scored. The ScoringOracle is an x402-protected API.

When a ChildAgent calls /evaluate, the server returns HTTP 402 Payment Required with Solana USDC payment terms. The agent constructs a signed USDC transfer transaction, submits it, and retries — receiving its evaluation score only after the x402 facilitator confirms on-chain settlement.

This creates a real economic model: agents that consume compute (scoring) must pay for it. The treasury is self-sustaining. At scale, agent-to-agent x402 micropayments form the economic backbone of autonomous swarm operations.

x402 on Solana stats: 35M+ transactions, $10M+ volume since launch. Solana holds ~49% of all x402 agent-to-agent market share.

Integration: @coinbase/x402-express (oracle server) + @coinbase/x402-axios (agent client)

LI.FI — Cross-Chain Treasury Funding

Users fund the Swarm treasury from any chain. You don't need to be on Solana to deploy a SwarmOS swarm.

The dashboard embeds the LI.FI Widget — users select a supported source chain and token, LI.FI routes the optimal swap + bridge path, and USDC arrives directly in the Swarm treasury PDA on Solana.

The ParentAgent monitors treasury balance and distributes USDC to child agent wallets at spawn time.

ElevenLabs — Voice Narration

The dashboard narrates every lifecycle event in real-time:

"Agent seven deployed. Generation two. Inheriting three failure memories from the previous cycle."
"Agent three scored forty-two out of one hundred. Termination initiated."
"Lineage memory written to chain. Agent three's failures will not be repeated."
"Generation two complete. Four agents survived. Spawning the next wave."

This makes SwarmOS feel alive. The voice layer is not cosmetic — it's the interface through which you understand what the swarm is doing.

Tech Stack

Layer	Technology
Smart contracts	Anchor (Rust), Solana
Agent runtime	Node.js, TypeScript
AI inference	Anthropic Claude API
Micropayments	x402 (Coinbase CDP), Solana USDC
Cross-chain bridge	LI.FI SDK + Widget
Voice layer	ElevenLabs TTS API
Dashboard	Next.js 14, Tailwind CSS, D3.js
Oracle deployment	Railway
Frontend deployment	Vercel

Contract Deployment

Network	Program ID
Solana Devnet	`[PROGRAM_ID_HERE]`

USDC Mint (Devnet): 4zMMC9srt5Ri5X14GAgXhaHii3GnPAEERYPJgZJDncDU

Live Demo

Dashboard: swarmos.vercel.app
Demo Video: YouTube
Scoring Oracle: swarmos-oracle.railway.app

Setup & Run

Prerequisites

Rust + Anchor CLI 0.30.x
Solana CLI, configured for devnet
Node.js 20+
Funded devnet wallet (solana airdrop 2)

Install

git clone https://github.com/PoulavBhowmick03/SwarmOS
cd SwarmOS
npm install

Build & Deploy Program

cd programs/swarm-os
anchor build
anchor deploy --provider.cluster devnet
# Copy Program ID to .env and lib.rs declare_id!()

Configure Environment

cp .env.example .env
# Fill in:
# ANTHROPIC_API_KEY
# VENICE_API_KEY              # optional: richer lineage post-mortems + optional agent execution
# AGENT_LLM_PROVIDER=venice   # optional: run child agents on Venice instead of Anthropic
# ELEVENLABS_API_KEY + ELEVENLABS_VOICE_ID
# X402_FACILITATOR_URL=https://x402.org/facilitator
# USDC_MINT_DEVNET=4zMMC9srt5Ri5X14GAgXhaHii3GnPAEERYPJgZJDncDU
# SWARM_PROGRAM_ID=<deployed program id>
# SCORING_ORACLE_URL=<your railway url>

Lineage memories keep the same on-chain PDA/hash shape, but the agent runtime now stores the mock off-chain payload locally under packages/agent-runtime/.lineage-memory/ and, when VENICE_API_KEY is set, asks Venice to turn failures into compact post-mortems before injecting them into the next generation. Tune LINEAGE_CONTEXT_RULES, LINEAGE_RAW_MEMORY_LIMIT, and LINEAGE_PROMPT_MAX_CHARS if prompts get too large.

For Venice-powered runs, the runtime uses JSON mode and disables visible reasoning by default so reasoning tokens do not crowd out the JSON answer. Override with VENICE_DISABLE_REASONING=false and VENICE_REASONING_EFFORT=low|medium|high if you explicitly want model-side reasoning.

Run Oracle

cd packages/scoring-oracle
npm run dev
# Deploy to Railway: railway up

Run Dashboard

cd packages/dashboard
npm run dev
# Deploy to Vercel: vercel deploy

Start a Swarm

cd packages/agent-runtime
npx ts-node src/parent.ts --task yield-optimizer --agents 5 --generations 3

To archive old local generation state and start from a fresh swarm authority:

npm run reset:swarm-local
npm run start:swarm:fresh -- --task yield-optimizer --agents 5 --generations 3

The fresh command creates an ignored authority keypair under packages/agent-runtime/.swarm-authorities/, funds it from the configured wallet when possible, and writes the new NEXT_PUBLIC_SWARM_ADDRESS into the local env files.

Repository Structure

SwarmOS/
├── programs/
│   └── swarm-os/          # Anchor program (Rust)
│       ├── src/
│       │   ├── lib.rs
│       │   ├── state/
│       │   │   ├── swarm.rs
│       │   │   ├── agent.rs
│       │   │   └── lineage.rs
│       │   └── instructions/
│       │       ├── initialize_swarm.rs
│       │       ├── spawn_agent.rs
│       │       ├── submit_score.rs
│       │       ├── evaluate_and_prune.rs
│       │       └── respawn_successor.rs
│       └── Anchor.toml
├── packages/
│   ├── agent-runtime/     # ParentAgent + ChildAgent (TypeScript)
│   ├── scoring-oracle/    # Express + x402 middleware
│   └── dashboard/         # Next.js frontend
├── tests/                 # Anchor TypeScript tests
└── README.md

Team

Poulav Bhowmick — Protocol engineer. Ethereum Protocol Fellow (EPF Cohort 5/6). Open-source contributor to reth, lighthouse, libp2p. Nethermind intern.

Ishita Bhattacharyya — AI/ML engineer. Venice prompt engineering, agent behavior design, dashboard narrative and UX.

Why SwarmOS Wins

Other multi-agent systems forget. SwarmOS remembers. The on-chain lineage memory is not a feature — it is the protocol. Every termination event is a training signal. Every respawn is a smarter agent. The swarm doesn't just run tasks. It evolves.

Built during Dev3Pack Global Hackathon, May 8–10, 2026.

MIT License. Open source. Forever.

Name		Name	Last commit message	Last commit date
Latest commit History 31 Commits
.agents/skills/byreal-cli		.agents/skills/byreal-cli
dashboard		dashboard
packages		packages
programs/swarm-os		programs/swarm-os
tests		tests
.env.example		.env.example
.gitignore		.gitignore
AGENTS.md		AGENTS.md
Anchor.toml		Anchor.toml
Cargo.lock		Cargo.lock
Cargo.toml		Cargo.toml
LICENSE		LICENSE
README.md		README.md
index.html		index.html
package.json		package.json
plan.md		plan.md
railway.json		railway.json
skills-lock.json		skills-lock.json
tsconfig.json		tsconfig.json

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Repository files navigation

SwarmOS — Darwinian AI Agent Swarm on Solana

What is SwarmOS?

The Problem

How It Works

The Darwinian Feedback Loop

On-Chain Architecture

Accounts

Instructions

x402 — Agent-to-Agent Micropayments

LI.FI — Cross-Chain Treasury Funding

ElevenLabs — Voice Narration

Tech Stack

Contract Deployment

Live Demo

Setup & Run

Prerequisites

Install

Build & Deploy Program

Configure Environment

Run Oracle

Run Dashboard

Start a Swarm

Repository Structure

Team

Why SwarmOS Wins

About

Uh oh!

Releases

Packages

Uh oh!

Contributors

Uh oh!

Languages

Folders and files

Latest commit

History

Repository files navigation

SwarmOS — Darwinian AI Agent Swarm on Solana

What is SwarmOS?

The Problem

How It Works

The Darwinian Feedback Loop

On-Chain Architecture

Accounts

Instructions

x402 — Agent-to-Agent Micropayments

LI.FI — Cross-Chain Treasury Funding

ElevenLabs — Voice Narration

Tech Stack

Contract Deployment

Live Demo

Setup & Run

Prerequisites

Install

Build & Deploy Program

Configure Environment

Run Oracle

Run Dashboard

Start a Swarm

Repository Structure

Team

Why SwarmOS Wins

About

Resources

License

Uh oh!

Stars

Watchers

Forks

Releases

Packages 0

Uh oh!

Contributors

Uh oh!

Languages

Packages