Cortal is an applied AI company building persistent caseworkers for financial operations. We’re building a future where a team can hand real, regulated work to software and still know, precisely and after the fact, what was done, what was held back, and why.
While model capability has advanced dramatically, the operational gap has not closed. A single agent is an unreliable part: fluent, fast, and confidently wrong often enough that no one can hand it the keys. The industry’s answer has been to add more model . longer contexts, better prompts, bigger evals. That improves the average case and does nothing for the case that matters, which is the one where the agent is sure and wrong. To close the gap, we build the layer around the model, a control loop that decides whether work may run, waits when it should, stops when it must, and keeps the evidence either way.
We are engineers and operators who have run production systems where being wrong is expensive: payments, underwriting, claims, and the compliance surface around them. The work is durable state, event-anchored timers, and calibration against real outcomes.
Reliable systems from unreliable parts
The loop is the product. The internet makes reliable connections over an unreliable network by checking, retrying, and refusing to pretend. Cortal does the same for agent work. Every proposed action passes hard rules, then authority, then an expected-value check, and returns one of three verdicts: go, wait, or no. Nothing reaches a customer or a system of record without one.
State outlives the process. Cases, timers, gates, and receipts live in durable storage, not in a context window or a long-running process. A worker can crash mid-case, a deploy can roll, a week can pass. The case resumes where it stopped, with its deadlines intact and its evidence attached.
Proof, not confidence
Confidence numbers are decoration until they are measured. If a system says it is eighty percent sure, it should be right about eighty percent of the time. We score that against what actually happened, per step type. When it sounds sure, it has to be right.
Receipts, not self-report. An agent saying it did something is not evidence that it did. Cortal records proof from tool reads, captured artifacts, and system observation. An unverifiable claim is treated as an unfinished job, not a completed one.
Autonomy is earned, never granted. Routine authority is unlocked one step type at a time, on measured outcomes, and is withdrawn the same way. There is no global switch that makes an agent trusted.
Solid foundations matter
Boring infrastructure, deliberately. Durable state, idempotent writes, and event-anchored timers are unglamorous, and they are most of the work. We would rather be correct on the tenth replay than impressive on the first run.
Recovery is a feature, not an incident response. Partial failure is the normal case in long-running work. Cortal is built so that an interrupted case, a duplicated event, and a late external callback all resolve to the same state.
The work stays where it lives. Cortal holds control state and references to evidence. It does not become a second system of record, and it does not ask an operations team to migrate before they can see value.
Learning by doing
We run it on ourselves first. Before any regulated workflow, Cortal runs our own operations end to end. If it cannot survive its own author’s week, it is not ready for someone else’s quarter.
Product and research co-designed. Deployment is how we learn what the control loop actually needs. Real cases keep us honest about which failures are common and which are merely interesting.
Measure what matters. Not tokens, not tasks attempted. Accepted work. The share of real jobs that finish in a result an operator can use, with the uncertain ones surfaced rather than buried.
Talk to us
We’re building Cortal for teams that run financial operations and want to hand real work to agents without losing control. If that is your desk, we would like to compare notes.