Runtime Labs
What we work on
We study how agents reason, use tools, and follow controls across fintech operations. Real work, such as ACH returns and settlements, keeps the research grounded.
Evals: graded on real cases
Test agents against real financial work and the policies behind it. Measure accuracy, supporting evidence, and whether the agent respects required approvals.
Routing: the smallest model that passes
Use the smallest model that meets each task’s quality bar. Routine checks and complex financial investigations need different depths of reasoning, with consistent context and controls.
Reliability: agents that stay up
Keep operations recoverable when a sandbox or provider fails. We test checkpoints, retries, and handoffs so work resumes with its context, evidence, and approval state intact.
Memory: every case compounds
Turn findings and decisions into reusable organizational context. Agents can draw on what teams have learned, with sources, permissions, and an audit trail attached.
Own a workflow, not a task
Across fintech operations, agents should carry work from intake through resolution. The workflows below show a few ways to measure a complete, verifiable outcome.
| Workflow | Outcome | Illustrative activity |
|---|---|---|
| Reconciliation | Breaks closed and days to close | |
| Bank RFIs | Replies before the bank deadline | |
| Disputes | Evidence accepted and win rate | |
| Underwriting | Decisions per analyst, with evidence | |
| Payment support | Escalations kept out of engineering | |
| Fraud monitoring | Alerts worked and false positives cleared | |
| Compliance | Alerts closed with an audit trail |
Give agents more responsibility, at your pace
Start with reviewed work and expand responsibility as agents prove reliable. You define the policies, limits, and approvals at every stage.
Start with completed work. Check whether the agent follows your process, uses the right evidence, and reaches the expected result.
Let agents investigate live work and prepare the next step. Your team reviews the evidence and approves the action.
Allow actions you’ve already approved, within your policies and limits. Keep sensitive decisions and money movement behind approval.
Agents carry approved work from intake to completion. Bring in a person when a decision needs judgment or additional approval.
The language of fintech operations
From payment standards to risk policies, financial work has its own context. We study how agents use that context accurately, while keeping organizational knowledge governed.
Field notes and research
Notes on model behavior, harness design, and what we learn running agents in real fintech workflows.
Field notes
01What is an agent harness?02What is a system of intelligence?03How to reduce AI agent token costs04How to reconcile payments with Claude Code and Codex05How to investigate payment declines06How to handle bank RFIs with agents07How to automate merchant underwriting08How to investigate fraud alerts09How to investigate compliance alertsOpen model research
Best open source models for payment reconciliationBest open source models for partner-bank RFIsBest open source models for transaction monitoringBest open source models for chargebacks and disputesBest open source models for merchant underwritingBest open source models for KYC and KYB checksBest open source models for AML and compliance investigationsBest open source models for payments customer supportAll open source models, rankedAbout Runtime
Runtime gives fintech teams one platform to build and run agents across payments, risk, compliance, and support. Runtime Labs is our applied AI work on making those agents reliable on real workflows. We’re based in San Francisco, with a team that has built payment and risk systems at fintechs and Fortune 500 companies.






Bring us your hardest workflow
We’ll build the first agent with your team and measure it against the cases you already resolved.







