Skip to content

AI

Agents and scale

The agent architecture, permission boundaries and application workflows behind concurrent AI activity.

AI-scale concurrency

Scale the number of AI jobs without putting the model inside consensus.

Research target

Engineering target

100,000–500,000 concurrent AI jobs

Engineering target; subject to benchmark validation.

This is not 500,000 finalized Layer-1 transactions per second.

What enables the target

  • 01Off-chain execution
  • 02Distributed provider mesh
  • 03Asynchronous jobs
  • 04Transaction lanes
  • 05Parallel execution
  • 06Low-contention state
  • 07Receipt batching
  • 08Merkle commitments
  • 09Settlement netting

Native finalized L1 throughput remains an engineering / benchmark target until reproducible multi-validator testing demonstrates it.

Agents that can request, verify, and settle

Give AI agents economic rails without giving them unlimited authority.

  • Research Agent

    • Request data
    • Run multiple models
    • Verify sources
    • Pay providers
  • Commerce Agent

    • Compare products / services
    • Request quotes
    • Execute approved purchases
  • Treasury Agent

    • Monitor balances
    • Execute approved treasury policies
    • Request AI analysis
  • Enterprise Workflow Agent

    • Invoke internal / private models
    • Trigger workflows
    • Record proof
  • Cross-Chain Agent

    • Receive an event on one chain
    • Request AI
    • Return the result elsewhere
  • Market / Portfolio Agent

    • Analyze market data
    • Produce recommendations
    • Execute only within explicit permissions and applicable law

Agent actions must be constrained by user authorization, smart-contract permissions, risk limits, custody rules, and applicable regulation.

Agents are not presented as having any unrestricted ability to trade securities.

Research, education and human expertise remain central to useful AI systems.Photo by Vitaly Gariev · Unsplash

Autonomy with boundaries

An AI agent should know what it may do — and what it may not do.

Planned
An agent intent passes through a policy engine. If allowed, it executes. If not, it is rejected or sent for human review.
Agent intentPolicy engineAllowed?
Yes → Execute
No → Reject / human review

Possible controls

  • Spending limit
  • Approved assets
  • Approved smart contracts
  • Approved destinations
  • Transaction-size limit
  • Daily limit
  • Model whitelist
  • Provider whitelist
  • Required human approval
  • Emergency pause
  • Expiration time
  • Jurisdiction / policy rules
Agent authorization framework — planned application-layer capabilityPlanned

An open market for intelligence

Models, compute, retrieval, verification, and agents can become composable services.

Planned

Participants

  • AI user
  • Agent developer
  • Model provider
  • GPU provider
  • RAG provider
  • Storage provider
  • Verifier
  • Specialist agent
  • Appchain
  • Validator

Service categories

  • Text inference
  • Image generation
  • Multimodal analysis
  • Code execution
  • Retrieval
  • Legal research support
  • Finance analysis
  • Enterprise automation
  • Agent-to-agent services

Legal and financial examples describe research and analysis tools. They do not imply licensed professional advice or securities execution is automatically permitted.

From AI output to deterministic action

Verified AI results can trigger programmable workflows.

  • Insurance

    AI

    AI assesses document evidence.

    Contract

    Contract applies an approved business rule.

  • Supply chain

    AI

    AI detects an anomaly.

    Contract

    Contract triggers an audit workflow.

  • Commerce

    AI

    Agent negotiates.

    Contract

    Contract enforces payment conditions.

  • Gaming

    AI

    AI generates outcome / asset metadata.

    Contract

    Contract records ownership.

  • Tokenized assets

    AI

    AI analyzes eligible data.

    Contract

    Contract records compliant workflow state.

AI should not have unchecked authority over irreversible smart-contract actions. Applications should define confidence thresholds, permissions, human review, and fallback rules.