We have created an architecture built on top of Harness.

This work presents Argus, a general agentic runtime built for systematic exploration of next‑generation autonomous research systems. Instead of merely enabling agents to reflect multiple times within a single task, Argus targets a critical question: can agents achieve continuous learning, adaptation and self‑evolution for long‑term autonomous operation without constant model‑weight updates, by leveraging persistent memory, external verification and accumulated experience?

Powered by persistent, verification‑guided and self‑evolving mechanisms, Argus stores validated knowledge, skills, verifiers, failure traces and decision insights across sessions. It makes autonomous decisions to persist, roll back or pivot based on new evidence, and runs unattended most of the time while reserving key decision points for human‑in‑the‑loop feedback.

Highly pluggable, extensible and deployable, Argus allows free replacement of models, tools, memory modules and workflows. Users can encapsulate domain‑specific expertise into verticals and deploy on cloud, local or private infrastructure. Rooted in Expert‑Agent Co‑evolution, experts set goals and evaluation criteria, while agents conduct exploration, execution and validation, enabling professionals and specialized agents to advance research collaboratively.