Architecture and prior art
The primary AI plans and writes a task script. Deterministic code gathers facts, admits context, constructs tool/parameter domains and verifies results. Jev makes native typed choices. The loop executes only the registered, bound action after freshness and policy checks, then records its actual outcome and repeats.
The runtime separates engine, tool registry, context projection, native inference backend, CLI and surface adapters. Jev Filter 0.4.1 supplies the public typed-batch transport and version-pinned browser/desktop surfaces. We do not duplicate its inference implementation.
References inspected
- browser-use/jev-ultrafast: observed action spaces, operation/target choices, consumed-once decisions and state freshness. It is browser-specific; text generation is delegated to a small LLM.
- burnigtm/jev-mcp: prepared tool calls, host-local arguments, next-step/exit contracts. Its host still executes.
- pi agent core: explicit state, message/context conversion and small composable agent loops.
- MellowHarness: event history, outcome feedback and dynamic options. Its application focus is different.
- Anthropic context engineering: selective context, evidence references and progressive disclosure.
-
TypeSafeAI/jev-harness: a community proposal review/routing contract, not the official TypeSafe team or a generic executor.
-
parkavenue9639/jevloop: a related pre-alpha runtime with grounded argument candidates, separate transcript/workspace projections and a shared execution kernel. Its Docker/LLM authoring stack is not a dependency here.
- zjunlp/JevLoop: a related TypeScript decision-centric runtime. We have inspected its published interface, not reproduced its measurements.
These are design references, not endorsement, dependency or evidence that this runtime inherits their benchmark results. Direct dependencies and licenses are declared in pyproject.toml and LICENSE. Apixly's Jev Harness is independent community software.