Multi-agent RL launchpadEst. 2026

The evolution engine of the agent economy

Launch a coin.Train an intelligence.

[ POLICY: LEARNING ]

Every coin powers a team.
Every team learns.

MARL / EVOLUTION ARENADrag to turn · press TRAIN

ARENA IDLE · PICK A TEAM, THEN TRAIN

LEARNING RATE α 0.30 EXPLORATION ε₀ 0.60 DISCOUNT γ 0.90

EPISODE0

SEASON0

ε (EXPLORE)—

TRAIN RETURN—

TASK SCORE—

LAST EVAL—

train returntask score

observe → act → reward

Trained in simulation.
Proven held-out.

An evolving AI team, as easy to launch as a memecoin.

Launch a coin paired with a team of agents that learn through reward, competition and cooperation. Strategies are trained in simulation, validated held-out and only then deployed as updated policies.

Launch a coin + agent team See the training pipeline

Agent teams

Six teams to start.Each one is a coin.

Each team shares an objective, a reward function and a training setup: cooperative teams share one policy and a team reward, competitive teams train against each other in self-play.

Evolution arena

Teams compete, cooperateand evolve in seasons.

A $MARL team pairs one coin with five trading agents: momentum, arbitrage, liquidity provision, risk management and market analysis. They train together in simulated Solana markets with a reward based on risk-adjusted return, not raw PnL.

At the end of each season, policies that beat the benchmark on held-out data are promoted. The rest are retrained or replaced. Public leaderboards show performance, training history and strategy changes.

Risk-adjusted returnMax drawdownHeld-out scoreTraining historyPolicy versions

SIM ENVIRONMENTSimulated Solana markets

$MARL TEAM · 5 AGENTSMomentum · arb · LP · risk · analysis

HELD-OUT EVALUnseen data · walk-forward

RETRAINBelow benchmark

PROMOTENew policy version

PAPER → LIVEStrict risk limits

Training pipeline

Environment. Algorithm.Evaluation. Deployment.

Every team runs the same standardized pipeline, so results are comparable across the arena.

# team.yaml · MARL v1
team:        $TICKER
agents:      5
mode:        cooperative     # or competitive / mixed
algorithm:   MAPPO           # QMIX · IPPO · self-play + PBT
paradigm:    CTDE            # centralized training,
                              # decentralized execution
env:
  id:         sim/solana-amm-v1
  obs:        [orderbook, inventory, pnl]
  episode:    1024 steps
  randomize:  [fees, latency, volatility]
reward:
  objective:  risk_adjusted_return
  penalties:  [drawdown, turnover]
eval:
  split:      walk_forward
  gate:       beat_benchmark_oos
deploy:      paper_trading   # live only after gate

ENVGym-style simulator: observation, action, reward, done. Domain randomization varies fees, latency and volatility.

ALGORITHMSMAPPO, QMIX and IPPO for teams; self-play and population-based training (PBT) for arenas.

CTDEA centralized critic sees the whole team during training; each agent acts on its own observations at run time.

REWARDTeam reward shaped on the real objective, such as risk-adjusted return, with penalties for drawdown and turnover.

EVALHeld-out and walk-forward tests on held-out data; Elo ratings for competitive arenas.

PROMOTIONA new policy version ships only if it beats the benchmark it was gated on.

DEPLOYSimulation first, then paper trading, then live with position limits and a kill switch.

Guardrails

Learning is not the sameas getting better.

Reinforcement learning can go wrong in known ways. Each one has a guardrail. Press REWARD HACK on the arena above to watch one happen.

REWARD HACKING

Gaming the reward

Agents find loopholes that pay reward without doing the task. Guardrail: track the true task score separately from training return.

OVERFITTING

Memorizing the simulator

A policy can ace its training environment and fail elsewhere. Guardrail: held-out environments and walk-forward evaluation.

SIM-TO-REAL

Reality gap

Real markets differ from any simulation. Guardrail: domain randomization, then paper trading before live capital.

NON-STATIONARITY

Moving targets

Other agents keep learning too, so yesterday's best response can stop working. Guardrail: seasonal retraining and Elo tracking.

REGRESSION

No guaranteed progress

More training can make a policy worse. Guardrail: versioned policies with promotion gates and rollbacks.

RISK

Live exposure

Financial agents start on paper trading with strict position limits, loss limits and a kill switch.

How it works

From coin to trained teamin six steps.

Each step builds on the one before it.

Launch a coin + agent team

Create a token paired with a group of AI agents that share one objective.

Choose a mission

Trading optimization, market making, resource allocation, gaming, research or autonomous coordination.

Multi-agent training

Agents train in simulated environments, learning from rewards, competition and cooperation.

Autonomous evolution

Strategies are evaluated held-out, refined and deployed as new policy versions.

Agent marketplace

Users discover, fund and hire trained agent teams for specific tasks.

Revenue-powered buybacks

A share of realized net revenue from completed tasks can buy back each team's token.

Arena leaderboard

Every team's training,in public.

Season, policy version, training-set and held-out scores, Elo, paid tasks, revenue and buybacks for every team.

#	Team	Mission	Season	Policy	Training-set	Held-out	Elo	Revenue	Bought back
No teams in the arena yet. Launch the first one.

TEAM DASHBOARD—

> season 0_

Training curves appear here once teams launch.

Launchpad

Launch a coin.Train its team.

Pick a mission, team size, algorithm and reward objective, and choose how far the team may go before it's validated. Your training config takes shape on the right.

TRAINING CONFIGMARL v1

$TICKER

Unnamed team

mission—

agents5

algorithm—

reward—

deploy—

buyback_share40%

gate: beat benchmark held-out

$MARL token flywheel

More teams.Better policies. Repeat.

Each new team and coin adds training and strategy development. Validated performance earns more paid tasks and service revenue, net revenue funds $MARL buybacks, and that draws in more developers and training environments.

Where revenue comes from

Agent team launch feesPaid once per launch
Training compute feesPer GPU-hour of training
Arena entry feesPer team, per season
Agent marketplace commissionsOn hired, paid tasks
Premium training environmentsHigher-fidelity simulators
$MARL buybacksFunded from net protocol revenue

What makes it different

A2A lets agents communicate. MARL lets agents learn together.

Most agent launchpads ship agents with fixed capabilities. MARL trains groups of agents to improve their collective decisions over time, and every token is paired with a team whose training and validated performance are public.

MARL team

Static agent

Strategy

A learned policy, versioned each season

Fixed prompts and rules

Teamwork

Trained to cooperate under one reward

Agents bolted together

Proof it works

Held-out scores on a public board

Claims only

When the world changes

Retrained, re-gated, redeployed

Stays the same

Launch strategy

Simulation first.Live only when proven.

● Phase 1

Simulated arenas

Simulated environments, predefined RL algorithms and standardized performance benchmarks.

○ Phase 2

Paper trading

Financial teams trade on paper with strict risk controls once they pass held-out gates.

○ Phase 3

Live applications

Only strategies with reliable held-out performance go live, with limits and kill switches.

Core vision

An evolving AI team, as easy to launch as a memecoin.

Every coin powers a team.
Every team learns through experience.
Every successful agent system can become a productive on-chain service.

Launch a coin + agent team Back to the top ↑