Agents — On-Device

On-Device Intelligent Agents

2025 — 2026

I work on intelligent agents whose core reasoning runs on the device, with the surrounding harness managing context, tools, permissions, durable state, verification, and recovery. The technical note below goes deeper into the architecture and system constraints behind this work.

An abstract local model core surrounded by layered control paths

Technical note Building Local Agents with Local Models A longer look at context budgets, harness design, reliable execution, and learning loops for agents that reason on the device. Read the article

Two system views

A local model loop surrounded by context policy, permissions, durable state, tools, validation, verification, observability, recovery, and evaluation — **Harness control plane.** Deterministic systems surround the model loop and connect it to users, tools, state, and the environment.

An execution path followed by trace replay, failure attribution, system or model repair, evaluation, and staged rollout — **Execution and attribution.** Actions are verified first; replay then identifies whether the harness or model owns a failure.

← Back to Projects