For most of the last three years, the industry has been building assistants. An assistant takes a task, produces an output, and hands it back to a person who decides what to do next. That loop is useful. It is also the wrong abstraction for the work that actually costs institutions money.
When I looked at where floors of specialists still sit in finance, semiconductors, energy, and pharma, I did not find tasks. I found lifecycles. A lifecycle is a multi-party process that starts with a signal, moves through a series of decisions and approvals, executes against an external counterparty, and ends with a record that a regulator can audit years later. ETF creation and redemption is a lifecycle. A letter of credit is a lifecycle. Batch release in a pharmaceutical plant is a lifecycle.
The defining feature of a lifecycle is that most of its cost lives in the exceptions. The happy path is already automated. What consumes human hours is the reconciliation break, the counterparty that did not confirm, the order that partially filled, the approval that needs a named person because a cap was breached. Software has always waited for a human to decide at those moments. My research thesis is that it no longer has to.
So we changed the unit of work. OpenEXA does not build an agent that does a task. We build infrastructure on which a swarm of small agents runs an entire lifecycle, from the first signal to the final ledger entry, with a human in the loop whenever a policy says there must be one and not otherwise.
The lifecycle test
Not every process qualifies. Early on, we wrote down six traits a process must have before we will touch it. It must be multi-party, regulated, structured enough to be machine-actionable, exception-heavy, gated by approvals or settlement, and bound to an audit trail. All six, or it isn't ours.
Those traits are not arbitrary. Each one maps to a layer in the system we built. Multi-party means we need authenticated connections to counterparties. Regulated means a policy engine must sit between decision and execution. Structured means an MCP server can expose the process as tools. Exception-heavy is where the agents earn their keep. Gated tells us where the council sits. Audit-bound tells us why the ledger is append-only and hash-chained.
If a process fails one of the six, the architecture has a piece with nothing to hold. That is why the list is a filter and not a wishlist.
Why finance first
We deliberately chose the smallest, lowest-risk lifecycle we could find as our proving ground: matching a fund's price to what it holds. ETF create and redeem is narrow, the rules are public, the counterparties confirm every action independently, and the downside of a mistake is bounded by position caps. It is the ideal environment to prove that thousands of agents can run a regulated process end to end without a person watching every step.
Trust is earned in the lowest-risk lifecycle first. Then it compounds. The next posts in this series describe how.