An AI agentic company
Infrastructure for agentic lifecycles.
OpenEXA builds the end-to-end infrastructure on which swarms of domain-specific AI agents run the world's high-value lifecycles.
Not tasks. Lifecycles.
The most valuable work in the economy is not a task an assistant can finish. It is a lifecycle: many parties, a regulator, structured data, exceptions, approvals, settlement, and an audit trail that has to survive scrutiny. Six industries share those six traits. One infrastructure runs them.
Eight layers. One autonomous stack.
Signal intelligence, a post-trained model layer, domain-specific agents, an execution council, routing, execution, a custom MCP server and an authentication layer — built as one system, so an agent can perceive, decide, act and prove without a human in the loop, and with one whenever you want.
Five to fifty thousand agents. Each does one job.
Every point on this screen is an agent. Signals funnel in; narrow agents predict, decide, approve, execute and settle down a vertical rail; rejected proposals leave at the gate; every settled transition is written to a hash-chained ledger.
ETF creation and redemption. Live. Ten for ten.
Our first lifecycle runs on live capital: more than ten basis points captured in every one of ten sessions, each fill confirmed against the executing broker's own records — with 90% of the traditional cost and risk gone. Three beta customers. Five lifecycle runs.
- Infrastructure for agentic lifecycles · signal → decision → execution → settlement → audit
- Swarm 5,000 – 50,000 domain-specific agents per lifecycle
- Lifecycle 01 ETF creation & redemption · live on real capital
- Proof >10 bps per session · ten consecutive sessions · broker-confirmed
- Economics 90% of cost & risk eliminated by agents
- Governance execution council · scoped permissions · hash-chained ledger
- Next six lifecycle classes · one infrastructure
Where work is a lifecycle, agents change the economics.
We don't build assistants. We build the infrastructure for agents to run entire lifecycles — the multi-party, regulated processes that today employ floors of specialists and still settle by exception. A lifecycle qualifies when it has all six traits.
The lifecycle testMulti-party orchestration
Banks, brokers, custodians, regulators, shippers, insurers — each holding a piece of the truth. Agents hold the whole sequence.
Regulatory oversight
Rules that decide what may happen, when, and by whom. Encoded as policy the execution council enforces on every proposal.
Structured data exchange
Messages, confirmations, filings and feeds in fixed formats. Machine-readable by design, machine-actionable at last.
Exception management
Breaks, discrepancies, partial fills, missing documents. Narrow agents resolve them; the council escalates what they can't.
Settlement or approvals
Value moves only after a gate: approve or reject, automatic or human. Nothing settles that didn't pass.
Audit-trail requirements
Every transition written to an append-only, hash-chained ledger — replayable by an auditor, a regulator, or you.
Financial markets & insurance
ETF creation and redemption, in production on live capital. Insurance and multi-asset settlement follow on the same rails.
International trade · letters of credit
Issuance, document checking, discrepancy handling and payment across issuing banks, shippers, insurers and beneficiaries.
Asset securitization · MBS / ABS
Pool assembly, tranche compliance, servicer reporting and investor distributions — monthly cycles run by exception today.
Wholesale energy settlement
Metering, scheduling, imbalance and multi-party settlement across system operators, traders and utilities.
Semiconductor manufacturing
Order-to-wafer lifecycles across foundries, packaging houses and logistics, with yield exceptions and export controls.
Pharmaceutical & aerospace
Regulated qualification, batch release and full traceability across suppliers, regulators and operators.
Eight layers. One boundary: intelligence decides, execution is deterministic.
Read top to bottom, from perception to permission. Each layer has one job and a hard edge — which is what lets thousands of narrow agents run at once without any of them becoming a black box.
Signal IntelligencePerceive
Proprietary AI and ML models for risk and prediction. For Lifecycle 01, the ETF create-and-redeem signal: NAV gaps, regimes, borrow and venue depth — continuously re-estimated.
- Risk model
- Price model
- Regime detection
Model LayerReason
A post-trained LLM tuned on the domain's documents, rules and exceptions — the substrate every agent reasons on, with tool use bounded by the layers beneath it.
- Post-trained LLM
- Domain corpus
- Bounded tool use
Domain-Specific AgentsAct
Proprietary, fully automated workflow agents — each built for a single, narrowly defined job inside the lifecycle. Five to fifty thousand of them make up a swarm.
- 5,000 – 50,000 per lifecycle
- One-click workflow creation
- Exception handling
Execution CouncilGovern
Arbitrates when agents disagree, enforces policy, and decides what passes the gate. Approve or reject, in milliseconds — with a human in the loop whenever the mandate says so.
- Approve / reject
- Policy engine
- Conflict arbitration
Routing LayerOptimize
Continually optimizes execution flow across venues and counterparties on live conditions and historical performance — the cheapest, fastest, safest path for every action.
- Venue scoring
- Cost of execution
- Smart routing
Execution LayerOrchestrate
Orchestrates the lifecycle end to end with deterministic order state — validated, submitted, partial, filled, reconciled — plus retries, idempotency and reconciliation against counterparty truth.
- State machine
- Idempotent
- Reconciliation
MCP ServerConnect
A custom Model Context Protocol server and toolset give agents authenticated connectivity to brokers, custodians, exchanges and clearing rails — deployed as secure, customer-scoped MCP servers.
- Custom tools
- Secure customer MCP servers
- Adapters
Authentication LayerPermit
Scoped, revocable permissions for every agent, tool and counterparty. Agents act inside limits you set; custody never moves; nothing has withdrawal rights.
- Scoped permissions
- Revocable
- Custody never moves
Trillions in 17,000 ETFs, created and redeemed by a swarm.
We chose the first lifecycle for its shape, not its glamour: super-low risk, high demand, a single agent type — and all six traits in full. Wall Street leaves the 4–8% band alone because human operations can't afford it at $0.0002 per dollar per day. Agents can.
- Real-time analysisData agents
- AI-ML predictionPrediction models
- Create / redeem decisionStrategy agents
human optional
- Clearing & settlementRail
- OTC & dark poolRail
- HFT automationRail
- AI/ML & strategyRail
- Dealer & brokerRail
The swarm at work. Tighten the council and watch what happens.
Every dot is an agent carrying one proposal through the six gates. At Approve the execution council rejects against a policy you can tighten; rejected proposals fall out of the lane, settled ones land in the ledger. One switch halts everything while open work drains.
Ten sessions. All above the floor. Confirmed by the broker.
Agents captured more than ten basis points in every session, in discount and premium regimes alike — the same swarm, the same gates, both sides of the NAV. Every fill is verified against the executing brokers' own confirmations.
90% of cost and risk, eliminated.
Agents don't carry inventory and don't wait for a counterparty, so more than 80% of the traditional stack disappears outright. What remains — technology and management — is compressed by 67% and 50%. It is the same result in every lifecycle class: the economics change because the operator changed.
The operating modelEvery agent acts on the record.
Autonomy is a dial, not a leap. Agents propose; the execution council approves or rejects against your policy; scoped permissions bound what any agent can touch; and every transition is written to a ledger you can replay.
- Council — approve / reject on every proposal, human optional
- Permissions — scoped, revocable; custody never moves
- Ledger — append-only, hash-chained, replayable
- Modes — automatic · with approval · manual, per lifecycle
- Halt — database-backed, available in every mode
Eight records from one proposal, each hashed with SHA-256 over its contents and the hash before it. Click any event or detail and type: that record and every one after it stop verifying, in red, because the chain no longer agrees with itself.
- Enable JavaScript to try this illustrative SHA-256 chain.
We publish the thinking behind the swarm.
Two research notes on why long-horizon autonomous work must be decomposed and how a model comes to follow one rulebook, then an eight-part series that walks the architecture gate by gate.
Compounding Error and the Case for Decomposition. A Research Note on Agent Architecture
Why long-horizon autonomous work must be decomposed into narrow agents behind a deterministic boundary.
6 min readSpecializing a Model to a Rulebook and Proving What It Did. A Research Note on Post-Training and Audit
How a general model comes to follow one rulebook, and how a hash-chained ledger proves what it did.
7 min readNot Tasks. Lifecycles. Why We Stopped Building Assistants
Why we stopped building assistants and changed the unit of work to the lifecycle.
3 min readBring us a lifecycle.
For institutions running high-value lifecycles, managers who want Lifecycle 01 on their assets, and partners building on agentic infrastructure.