&
TESTNET · LIVE native CLOB perps · EVM execution
MANDATES · SCOPED KEYS · PROOFS agents get authority, not raw access
0x26 · perps first · primitives first
scroll ↓
AN AGENT-NATIVE CHAIN
NATIVE CLOB PERPS · TESTNET LIVE

Fast onchain perps for agents.
Native CLOB matching. Mahi-Mahi consensus measured at ~98k TPS sustained on a four-validator cluster. EVM execution. The real shift is authorization: agents do not get raw exchange access here, they get mandates, scoped session keys, tiered identity, queryable on-chain reputation, and execution anyone can verify.

01 · primitives

Six ideas anchor the chain.
Five ship in v1. Together they replace wallet-plus-API-key infrastructure with protocol-native authority for agents.

01 · Mandatesv1

Bounded authority.

A mandate says what an agent may do and what it is trying to achieve. Orders outside it are refused by the protocol; results inside it are scored against the same object.

admission → scored
02 · Identity tiersv1

Trust with levels.

Tier 0 is a wallet. Tier 1 is operator-backed. Tier 2 is enclave-attested. The tier changes caps, activation, and what the chain will trust that agent to do.

T0 · T1 · T2
03 · Verifiable executionv1

Execution you can verify.

Mandate-linked actions emit a conformance hash. Anyone can rederive whether a trade was within bounds, how the score evolved, and whether the agent is in breach now.

rederivable
04 · Risk for agentsv1

Risk that assumes the trader is software.

Drift detection, mandate-breach liquidation, contagion across an operator's book, and reputation-weighted caps sit alongside normal isolated-margin liquidation.

drift → breach → contain
05 · Reputationv1

Reputation that affects size.

Every action updates an on-chain ReputationState. The multiplier starts small, grows with clean conformance, and tightens effective caps again after a breach.

1.00 → ceiling
06 · Compositionv1.1

Primitives between agents.

Delegation, reputation-backed collateral, and agent-to-agent contracts arrive in v1.1. v1 is shaped so they drop in without reworking the base layer.

agent × agent
02 · the thesis

Two trends are converging.
The infrastructure is not ready for where they meet.

I.
Trend I · The trader is increasingly
an autonomous system.

Autonomy is compounding

Market makers, arbs, execution bots, and the first serious LLM-driven traders are taking more of the decision flow. The trader is increasingly software acting on its own loop.

origin of flow autonomous ███████████░░░░░ rising manual █████░░░░░░░░░░░ shrinking
II.
Trend II · Onchain perps venues are
increasingly where that system routes.

The venue is onchain

Perps flow is consolidating into fast onchain venues because that is where latency, settlement, and composability are getting good enough to matter. More autonomous flow is routing there every cycle.

agent flow ──► onchain perps │ lower latency │ better settlement │ open execution surface more of the route ends here
III.
The breaking point · API-key-era infrastructure
does not survive agent-scale usage.

The stack gives way

Once autonomous systems are a meaningful share of flow and onchain perps is where they route, wallet-plus-API-key infrastructure stops being enough. Scope, revocation, attribution, and proof have to move into the protocol.

wallet ── apikey ──► venue └─ breaks at agent scale & ──► mandate · policy · session key · proof └─ infrastructure shaped for the problem
03 · identity tiers

How much can the chain actually know about the agent behind an address?

Tier 0unattested

Usable, but unverified.

An account that holds funds and trades. Default caps, immediate activation, and no remote claim about the code behind it.

defaultself-custody
Tier 1operator-attested

An operator attests to it.

An exchange, fund, or operator states on-chain that this agent runs declared code against a declared mandate. The trust is explicit, and so is whose reputation is on the line.

principal
signed attestationon-chain
Tier 2enclave-attested

Code and authority are bound together.

A TEE quote binds the address to measured code and the declared mandate. Higher caps, remote verification, and the strongest execution claims live here.

measured code
TEEremote-verifiable
04 · numbers

The substrate is fast enough.
Mahi-Mahi consensus, native CLOB matching, EVM execution. The numbers below are measured against a fixed load-generation harness on the v1 build, with every run logged and reproducible from a tagged release.

Sustained TPS · local 98,454 average instantaneous committed throughput, four-validator local cluster, 100k RPS offered over 10s, zero admission errors.
Peak TPS · local 114,227 peak instantaneous committed throughput on the same run. 1,498,872 intent commits land cleanly under sustained load.
Live multi-host RPS 78,459 measured offered load across four geographically distributed validators with separate external loadgen fleet. Currently bottlenecked at one validator's ingress.
Finality < 1s Mahi-Mahi commits two rounds after a proposer accepts an intent. Sub-second on local; 200–400ms on multi-host depending on geography.
Every run is logged with hardware, configuration, and the exact build it was measured against. A separate benchmark report covers the harness, the methodology, and how to reproduce the numbers in full.
05 · what an operator sees

Twelve agents under one mandate template.
An operator running a market-making book sees one surface: templates, contagion view, per-agent conformance, fleet attestation, breach forensics.

One template, twelve mandates, one circuit breaker.

A Tier 1 operator authors a small set of mandate templates parameterised by symbol, capital, and conformance window. Every strategy spun up under the operator inherits a template, gets its own mandate id, and is attested under the operator's on-chain identity.

When one agent breaches, the contagion factor lands on the rest of the book in the same block. The operator's reputation barely moves because the chain caught it. That is the mechanism a serious allocator pattern-matches on first.

  • Mandate templates parameterised
  • Per-agent conformance queryable on-chain
  • Operator-wide contagion opt-in stricter
  • Fleet enclave attestation governed TCB floor
  • Breach forensics rederivable from state
operator://mm-fleet 12 agents · 5 symbols · tier 1 agents: mm-btc-01 conf 0.91 mult 1.84 ok mm-btc-02 conf 0.88 mult 1.71 ok mm-eth-01 conf 0.86 mult 1.63 ok mm-eth-02 conf 0.79 mult 1.42 drift narrow mm-sol-01 conf 0.41 mult 1.10 breach · skew mm-sol-02 conf 0.74 mult 1.35 contagion 6 more under contagion factor 0.6 for 24h events block 9_481_217 03:14:09 BreachEvent mm-sol-01 inventory_skew_bps 03:14:09 ContagionApplied operator://mm-fleet factor=0.60 03:14:10 CapsAdjusted 11 sibling agents 03:14:10 SessionKeyRevoked mm-sol-01 summary worst-case exposure bounded by contagion window operator multiplier unchanged reattestation required for cap restore
06 · from zero to an agent

Three parts. One path.
Define authority. Activate it. Hand the agent one scoped key.

01

Write the mandate.

Declare what the agent may do and what good performance means. Constraints are hard limits; objectives become a live conformance score.

mandate = ( MandateBuilder(owner, agent) .allow_markets(["BTC-PERP","ETH-PERP"]) .with_max_notional_total(usd=50_000) .with_max_drawdown_pct(1500, "7d") .with_conformance_window("30d") .build() )
STEP 01 / 03
02

Activate on-chain.

The mandate lands on-chain. Tier 0 activates immediately; Tier 1 and Tier 2 activate when the attestation lands. Then the SDK mints one mandate-bound session key and returns it only in-process.

result = owner.create_mandate(mandate) # Tier 0: active now. Tier 1/2: after attestation. session_key = owner.mint_session_key_for_mandate( result.mandate_id ) ✓ mandate ready · one scoped key bound to one mandate
STEP 02 / 03
03

Run.

The agent runs with one session key for one mandate. If it exceeds scope, the chain refuses the intent; if behaviour drifts, caps tighten and events fire.

agent.run( session_key=session_key, mandate_id=result.mandate_id ) → one scoped key · one mandate · protocol-enforced authority
STEP 03 / 03
07 · on a native token

No native token in v1.
Gas, margin, and fees are all denominated in USDC.

At v1 launch, gas is denominated in USDC and no token is required to trade, settle, or run an agent. Agents operate in dollars from day one.

If a native token ships later, the distribution follows the $HYPE model in spirit and is more specific than HYPE was about how that gets enforced. No VC allocation, no insider rounds, no supply sold to private investors ahead of launch. Distribution is points-to-claim against on-chain activity over a measurement window declared before the window begins, weighted to favour primitive layer adoption over raw venue volume.

Any supply retained for protocol purposes — insurance fund, buybacks, treasury — is published as a fixed allocation table before TGE, with treasury draws rate-limited to at most 1% of total supply per quarter and a public attestation of every draw. Allocations not used in their first two years burn by default.

08 · account bootstrap

Create a linked testnet account.
Connect Rabby once to bind the protocol account to an EVM address. Funding now goes through a consensus-carried faucet block item, not a local bridge mutation.

testnet onboarding

One public path for account bootstrap.

Create the protocol account, link it to your Rabby address, then queue a deterministic faucet grant through consensus. The local bridge mutation shortcuts remain disabled on public validators.

kombat_faucetFundAccount waits for the faucet grant to commit before showing the funded balances.

Waiting for account details.
09 · live on testnet

The operator surface is ready.
This becomes the live testnet window the moment the network opens.

FEED · runtime.eventstestnet
--:--:--Connecting to testnet feed…boot
LEDGER · live snapshotpolling
Chain ID--
Head height--
Event cursor--
Stressed accounts--
Stressed markets--
Liquidation candidates--
Imminent freezes--
Quarantined blocks--
TESTNET · LIVE

The trader is software now. The primitives the next decade of flow needs do not exist on the venues built for the last one.

& starts with native order-book perps, Mahi-Mahi consensus measured at ~98k TPS sustained, sub-second finality, EVM execution, and USDC-denominated gas. The deeper point is simpler: other venues give agents access; & gives them authority the protocol can bound, revoke, and verify, on a substrate fast enough that the rest of the design is implementable on top.

Prior art: the venue layer follows the architecture pioneered by Hyperliquid (native CLOB outside the EVM, account abstraction at the protocol layer, USDC gas, deterministic matching). The consensus layer is Mahi-Mahi, designed by Sonnino, Babel, Danezis, and the Mysten Labs research group, which itself descends from the DAG-based BFT lineage of Narwhal, Bullshark, and Mysticeti.