August 2026 Summaries

Arcee AI introduces Nac, an Apache-2.0 open-source agent harness for complex, long-running tasks that separates temporary worker context from persistent task state to reduce context degradation. Based on a thread-and-episode architecture, Nac uses a non-executing orchestrator to plan and dispatch bounded tasks to fresh worker processes, which return concise episodes containing durable handoffs, results, and relevant artifacts while their full execution contexts are discarded. Threads retain ordered episode histories, and selected episodes can be routed across threads as dependencies, enabling parallel execution through validated acyclic task graphs while preserving synchronization points and failure reporting. Nac is positioned as an inference runtime rather than merely a workflow tool because it governs context construction, scheduling, tool-driven effects, persistent state, capabilities, failure behavior, and task completion. It is intended for decomposable work such as research reproduction, large code migrations, parallel code changes, reviews, experiments, and infrastructure tasks, while simpler single-session tasks may be better handled directly. Nac also includes an MCP server that allows interactive agents to launch, monitor, and steer background jobs, supporting a model in which human-facing agents manage longer-running computations.

Arcee AI has launched a beta expansion of its Open Models API, extending beyond its Trinity family to offer several frontier open models for demanding, long-horizon agent workloads. The initial catalog includes Trinity-Large-Thinking, DeepSeek-V4-Pro and V4-Flash-Latest, GLM-5.2, Kimi-K3, and Thinking Machines’ Inkling-Small, with token-based pricing ranging from $0.14 per million input tokens for DeepSeek-V4-Flash-Latest to $15 per million output tokens for Kimi-K3. Arcee says the broader selection gives developers and enterprises more flexibility while generating usage insights that can inform future Trinity model development, reflecting its view that different workloads require different models. The release coincides with the open-sourcing of nac, Arcee’s internal harness for long-running agent tasks, which is expected to help improve API routing, serving, and model support over time. New users who register and add a credit card receive $5 in API credits to test the catalog and nac.