Platform

The agentic stack. Eight layers, one autonomous lifecycle.

OpenEXA separates perception from decision from execution, binds them with authenticated connectivity, and writes every action to a log that cannot be edited. Built once. Deployed per lifecycle.

Live · Lifecycle 01 · ETF create & redeem
01ArchitectureCore infrastructure components · read top to bottom

Perceive. Reason. Act. Govern. Optimize. Orchestrate. Connect. Permit.

Each layer has one job and a hard boundary. That is what lets the system run tens of thousands of narrow agents at once without any of them becoming a black box — and what lets a new lifecycle reuse everything below the agents.

01

Signal IntelligencePerceive

Proprietary AI and ML models for risk and prediction. In Lifecycle 01 the create-and-redeem signal watches every ETF against its NAV, basket, borrow and venue depth, and captures short-lived discrepancies when execution latency is controlled and risk is properly managed. It never sleeps and never anchors.

  • Risk model
  • Price model
  • Arb model
  • Regime detection
  • Continuous re-estimation
02

Model LayerReason

A post-trained LLM tuned on the lifecycle's documents, rules, message formats and exception history — the substrate every domain agent reasons on. Tool use is bounded by the layers beneath it: the model can propose anything and execute nothing on its own.

  • Post-trained LLM
  • Domain corpus
  • Bounded tool use
  • Evaluated per release
03

Domain-Specific AgentsAct

Proprietary, fully automated, domain-specific workflow agents. A one-click workflow creator builds custom execution workflows tailored to specific market conditions and regulatory requirements. Each agent does one narrow job — read a feed, score a gap, size an order, check a limit, match a fill — and the swarm does the rest.

  • 5,000 – 50,000 per lifecycle
  • One-click workflow creation
  • Exception management
  • Master–copy replication
04

Execution CouncilGovern

The gate every proposal passes. The Council arbitrates when agents disagree, enforces policy and mandate limits, and returns approve or reject in milliseconds — or hands the decision to a human when the operating mode says so. Nothing reaches execution that did not pass.

  • Approve / reject
  • Policy engine
  • Conflict arbitration
  • Human-in-the-loop optional
05

Routing LayerOptimize

Continually optimizes execution flow for an efficient lifecycle. Routes each approved action across brokers, OTC desks, dark pools, exchanges and clearing rails on real-time conditions and historical performance — the cheapest, fastest, safest path for that action, right now.

  • Smart order routing
  • Venue scoring
  • Cost of execution
  • Latency budget
06

Execution LayerOrchestrate

Orchestrates the lifecycle end to end with deterministic state — validated, submitted, partial, filled, reconciled — plus retries, idempotency keys and reconciliation against counterparty truth. The execution layer stays algorithmic so that the intelligence above it can be audited against it.

  • Order-state machine
  • Idempotent submission
  • Retries & backoff
  • Reconciliation
07

MCP ServerConnect

A custom Model Context Protocol server and toolset deploy tools to the agents and provide authenticated connectivity to brokers, transfer agents, custodians, execution desks, exchanges and clearing. Deployed as secure, customer-scoped MCP servers — your rails, your context, nothing shared.

  • Custom tools
  • Secure customer MCP servers
  • Broker & custodian adapters
  • Real-time data pipe
08

Authentication LayerPermit

Scoped, revocable permissions for every agent, tool and counterparty — exactly the permissions a workflow needs, and nothing more. Agents act inside limits you set. Custody never moves. Nothing in the system has withdrawal rights.

  • Scoped API permissions
  • Revocable per agent
  • Never withdrawal rights
  • Role-based access
02One boundaryAdaptive intelligence vs. deterministic execution

Intelligence decides what, when and how much. Execution stays deterministic.

DeterministicLayers 05 – 08

Rules, adapters, order state.

  • Deterministic action rules with explicit pre-conditions
  • Adapters with retries, backoff and idempotency
  • State machine: validated → submitted → filled → reconciled
  • Reconciliation against counterparty truth
  • Market-hours gates and instrument-level constraints
AdaptiveLayers 01 – 04

Models, agents, council.

  • Signal selection and parameter tuning per regime
  • Risk-adjusted strategy rotation under portfolio-level constraints
  • Anomaly detection across signals, venues and counterparties
  • Arbitration between agents when their proposals conflict
  • Only profitable, policy-compliant actions pass to execution
03The swarm5,000 – 50,000 agents per lifecycle

Thousands of narrow agents beat one wide one.

A lifecycle is decomposed into hundreds of small, verifiable jobs. Each job gets an agent; each agent gets exactly the tools and permissions that job needs; the Council arbitrates. Nothing important depends on a single model's judgement — and every judgement is on the record.

The master–copy agent model. One master agent is proven in production, then copied and specialized — a copy per asset, venue or customer — all reporting to the same Council, all sharing the same ledger.

0k
Agents at scale
Elastic from 5,000 to 50,000 per lifecycle.
8
Layers
From perception to permission.
3
Operating modes
Automatic · with approval · manual.
1
Ledger
Append-only. Hash-chained. Replayable.
A-01Signal agents

Watch

Real-time analysis across prices, NAV, baskets, borrow, venue depth — and, in other lifecycles, documents, meters and filings.

A-02Strategy agents

Decide

Create or redeem, size and timing — proposed under risk, policy and portfolio constraints, then submitted to the Council.

A-03Execution agents

Act

Route approved actions to clearing, OTC, dark-pool and HFT rails and dealer-brokers through the routing and execution layers.

A-04Reconciliation agents

Prove

Match fills to counterparty truth, raise discrepancies, and close the loop in the hash-chained ledger.

04Execution pipelineFrom validation through reconciliation

Every action passes three gates. Every gate leaves a record.

Institution-grade controls for autonomous execution — designed so that an operator, an auditor or a regulator can replay any action, at any step, and see exactly why it moved.

  1. P-01

    Validate and risk-check

    Proposals pass schema checks, policy rules, position and notional caps, and market-hours gates before they exist as orders at all. The Council signs the result.

    schema ✓ · policy ✓ · caps ✓ · council ✓
  2. P-02

    Submit and confirm

    Submission tracks fills, partial fills and retries with idempotency keys, so a disrupted connection can never produce a duplicate action.

    submitted → partial → filled
  3. P-03

    Reconcile and report

    Internal records are matched against counterparty truth — positions, cash, NAV, documents — with discrepancy alerts raised to operators and written to the ledger.

    internal ≡ counterparty · alerts on Δ
05DeploymentSecure customer MCP servers · real-time data pipe

Your rails. Your context. Your permissions.

Every customer runs on a dedicated, secure MCP server: the agents that act for you see only your tools, your data and the permissions you grant — and can lose them in one call.

D-01Isolation

Dedicated MCP server

A customer-scoped Model Context Protocol server with its own toolset and adapters. No shared context between customers, ever.

D-02Data

Productionized real-time data pipe

Prices, NAVs, baskets, confirmations and reference data delivered to the agents with latency budgets and provenance on every record.

D-03Control

Permissions you can revoke

Scoped API permissions per agent and tool, granted through the authentication layer and revocable instantly. Custody stays at your broker.

06Operating modesAutonomy is a dial · per lifecycle, per mandate

As autonomous as your mandate allows.

Each lifecycle runs in one of three modes. Switch modes at any time; the emergency halt is database-backed and works in all of them.

ModeWho approvesLatencyTypical mandate
Fully automaticAgents execute within policyExecution CouncilmsLow-risk single-agent lifecycles
Auto-with-approvalAgents propose, humans confirmOperator or delegateseconds – minutesNew lifecycles, new strategies, larger tickets
ManualAgents recommend onlyOperatordiscretionaryEvaluation and onboarding
Emergency haltAvailable in every modeAny operatorimmediateCounterparty or API disruption, market events
07The marketplacePhase 03 · platform strategy

Built as a platform others deploy agents on — not a single product.

The same eight layers that run OpenEXA's first lifecycle are designed for third parties to deploy their own agents, lifecycles and controls — once trust and liquidity have been built in production.

Supply sideManagers · builders

Publish agents and lifecycles.

  • Workflow creation and instrument-level constraints
  • Streamlined approval workflows through the Council
  • Governance management and role-based access
  • Sophisticated models, real-time feeds and algorithms — only compliant actions pass
Demand sideInstitutions · investors

Allocate with full visibility.

  • Lifecycle and pool selection with strategy-level performance
  • Transaction visibility for execution-quality monitoring
  • A secure sandbox: agents execute, monitor and enforce risk controls
  • Custody stays at your broker throughout
Access

See the stack run on live capital.

Request a walkthrough of Lifecycle 01 — the swarm, the Council, the control plane and the ledger — on real trades.