nac: an open-source agent harness that throws away its own context

Arcee AI open-sourced nac, an agent harness for long-running tasks, under Apache 2.0. Arcee describes a central orchestrator that plans work and dispatches it to isolated workers. Each worker runs in a clean context, makes its edits, and returns a concise summary called an episode. The worker process then exits and its raw context is discarded, while the episode persists. Arcee says threads can pass episodes to each other to share context without carrying historical noise. Nac also ships with an MCP server, so interactive coding agents such as Claude Code or Codex can delegate long-running background work to it.

Arcee launched an open models API beta the same day, hosting models beyond its Trinity line, including DeepSeek-V4 Flash and Pro, Thinking Machines' Inkling Small, Z.ai's GLM-5.2, and Moonshot AI's Kimi K3. Arcee's stated reason is that long-horizon tasks need different model strengths at different steps. It also says neither it nor its partners monitor or train on API outputs, a claim from the post that was not independently checked.