Developers
We do not publish our stack on a marketing page. We share the patterns beneath it — the mechanics that determine whether an agent is safe to put in front of live customers and real money. Four of them carry most of the weight.
The patterns
An agent that decides, routes, and acts in a single step is impossible to debug when something fails. Felesh keeps the three distinct: what to do, where the work goes, and what touches the outside world. Each boundary is a point you can inspect, gate behind an operator, or halt.
Every organization runs in its own isolated environment — separate identities, separate data stores, separate secrets. No organization can see or reach another, and that boundary is enforced by design, not by convention. You operate your agents inside it, with data-residency options you can point to.
Nothing reaches production by hand. Configuration and agent behavior are versioned with immutable history, applied through a single reviewed path, and rolled back exactly the way they deployed. When you ask what changed, who approved it, and how to undo it, there is always a specific answer.
Work moves as events rather than long-held connections blocking on a result. A signal arrives, it is recorded, and the system reacts. This allows a long-running operation to survive a restart, retry without acting twice, and resume exactly where it left off. Capacity scales by adding workers, not by holding connections open.
The glass box
Every action an agent takes leaves a receipt — what it saw, why it acted, what it did, and under which autonomy. Nothing happens off the record.
Felesh runs live in financial services — environments where "the agent did something we cannot explain" is an unacceptable answer. The four patterns make this workable. Per-organization isolation keeps each institution's data to itself. Decoupling the decision from the act means an operator can sit at the gate before anything reaches a customer or a record. Auditable, reversible change — plus a receipt behind every action carrying the reasoning, the source, and the timestamp — means any step can be traced and undone. And because execution is event-driven, a case that runs for days does not lose its place when a process restarts. None of this is bolted on for compliance; it is simply how the platform operates.
This page stays at the level of architectural patterns by design. The specifics — how isolation is strictly enforced, how the decide-route-act boundary is implemented, how change is versioned, and the data layers beneath it — are best discussed directly, under NDA, with the engineers who built them. If you are the architect evaluating Felesh for your organization, that is the conversation worth having.