Research

Papers, preprints, and technical reports.

We publish formal research on the problems we have to solve to ship sovereign AI systems in production: evaluation methodology under compliance constraints, agent architectures with persistent memory, and the systems work that makes both possible.

• 2026-06-09

  The Model–Harness Configuration as the Unit of Agentic Capability

  Progress in autonomous LLM agents is conventionally attributed to the model: "model M scores X% on benchmark Y." Yet the same model embedded in different runtimes — different context management, tool surfaces, orchestration, and verification — succeeds or fails on the same task, so model-level attribution obscures a large and controllable source of performance variance. This paper argues that the appropriate unit of analysis for agentic capability is not the bare model but the model–harness configuration: the model together with the runtime that governs what it perceives, what it may do, what state survives across turns, what it may not do, how its work is evaluated, and how it recovers from failure. We define the harness and distinguish capability-of-the-model from capability-of-the-configuration; give an operating-systems-grounded taxonomy of harness components, each with its OS analog and the design question it poses; and synthesize empirical evidence that scaffolding shifts end-to-end task success substantially with the model held fixed. We draw out implications for benchmarking (configuration-level reporting), for open-weight competitiveness (a "build to delete" account in which weaker models need more scaffolding and stronger models need less), and for governance, where a runtime-enforced, logged control is auditable evidence while a prompt instruction is only behavior.

  Bert Colemont

• 2026-04-29

  Capability-Scoped Runtimes for Desktop Agents: Risk-Gated Execution, Durable Tasks, and Reversibility-Aware Containment

  Desktop agents are advancing on planning and perception while remaining fragile on execution. We argue this is a runtime problem, not a planner problem, and present pan-agent — a managed desktop-agent runtime that composes three load-bearing mechanisms: a pre-execution risk-gated classifier that intercepts dangerous tool calls, a durable taskrunner that survives crashes and zombie processes, and a reversibility-aware containment layer that records each side-effect with a typed receipt and wires it to per-tool reversers backed by capability-probed filesystem snapshots. The contribution is framed around a taxonomy that classifies every action as local-reversible, runtime-compensable, or externally-irreversible, and a four-experiment pre-registered evaluation plan over OSWorld, OS-Harm, RedTeamCUA, and a long-horizon crash generator. Implementation is open source — approximately 32 kLOC of Go, MIT-licensed.

  Euraika Labs Research Group