Research
What we are figuring out.
Open questions rather than finished papers. Some of this feeds directly into the agents we build and some of it is just interesting. None of it is written up yet, so there is nothing to download. Where a question has already changed the product, the abstract says which part.
AI Agent Reliability in Production
Most AI agent projects fail before reaching production. We are collecting the failure patterns, from orchestration breakdowns to cascading errors, and the architectural decisions that separate an agent that demos well from one that survives a real week. Written up as we go rather than from a finished study.
The Real Cost of Manual Operations
What repetitive manual work actually costs a business once error rates, turnover, and opportunity cost are counted. The question we care about is the one that decides against us as often as for us: when is answering by hand still the right call.
Agent Safety and Governance Patterns
How to build agents that know their limits. We document the patterns we use for confidence thresholds, human escalation, approval workflows, audit trails, and rollback mechanisms. Written for teams that need agents in production but cannot afford the risk of an uncontrolled system.
Internet-Native Payment Protocol (x402)
The HTTP 402 status code was reserved for "Payment Required" but never implemented. The x402 protocol finally puts it to use. We are researching how this open standard enables AI agents and automated services to pay for APIs, compute, and data instantly with stablecoins. Our focus is on practical integration patterns and what this means for building autonomous services.
Federated Communication Infrastructure
Every business runs on chat and almost none of them own it. We are studying federated communication protocols as an alternative to centralised platforms. The premise: an organisation should be able to host its own communication, reach other organisations without either of them joining the same platform, and keep end-to-end encryption while doing it. The first thing to come out of this is shipped rather than written up. An agent can be attached to your own Matrix homeserver as an application service, which means the messages stay on infrastructure you run.