September 2026 Summaries

HarnessRouter is presented as a unified runtime and control plane for agent harnesses, allowing products to run tools such as Codex, Claude Code, Hermes, Gemini CLI, and others through a common task-based API that manages sandboxed execution, streaming progress, sessions, files, artifacts, metering, and routing. Available as a managed cloud service and an Apache 2.0 self-hosted Community Edition, it implements the draft Unified Harness Protocol (UHP), an open standard defining compatible APIs, event streams, error handling, conformance classes, and a 64-check test suite. Its design treats the harness—the system coordinating tools, permissions, recovery, and workspaces—rather than the underlying model as the key interchangeable infrastructure layer, enabling configured agents to separate product-specific instructions and access policies from vendor-specific backends. The platform uses OpenAI Responses-style requests and Server-Sent Events, emphasizes idempotent task execution and persistent sessions, and advises applications to keep credentials and authorization server-side while treating stored task results as authoritative over live streams. The Community Edition packages a console, gateway, and runner in Docker with SQLite storage, per-session operating-system isolation, configurable provider credentials, and portability of harness configurations to the cloud. Benchmark results cited by HarnessRouter suggest that harness and model choices can substantially affect cost and latency, supporting its argument for experimentation and routing across configurations, while its roadmap signals expansion from coding workflows toward document, spreadsheet, presentation, image, and video deliverables.

Epsilla outlines an agent-native operating model in which AI systems are treated as organizational colleagues that complete defined tasks and produce verifiable artifacts, rather than chat tools that generate drafts for humans to finish. Its approach relies on HarnessRouter as a runtime that can route work among different agent harnesses and models, emphasizing that the most effective and economical configuration varies by task and over time. The model assigns agents roles and inherited access controls, places them in existing workspaces, and evaluates their activity through records of changed documents, tasks, code, or other artifacts rather than self-reported descriptions. Epsilla proposes introducing automation gradually, beginning with low-risk internal research and content work, then reviewable engineering changes, and finally customer-facing tasks that initially remain drafts until they meet performance gates. It measures cost per successful task, tail latency, human acceptance without edits, and configuration churn, while reserving irreversible actions such as deployments, payments, merges, and sends for named human approvers. The broader roadmap presents AI maturity as an incremental process in which each task class earns greater autonomy through evidence, portability is continually tested, and agent configurations are promoted or replaced based on traced operational results.