Arcee AI open-sources NAC, an agent harness built to fight context rot

Arcee AI has open-sourced NAC, an Apache 2.0-licensed agent harness built to keep long-running, parallel AI agent tasks from succumbing to “context rot,” the company said, publishing the code on GitHub on August 13, 2026.

Context rot — the performance degradation large language models show as a task grows long and their context window fills with history — is a central obstacle to reliable autonomous agents. NAC’s design tries to route around it with a thread-and-episode architecture.

Arcee AI said a central orchestrator plans and dispatches work to specialized worker threads but cannot itself execute commands or edit files. Workers return “episodes” — structured summaries of completed work — that persist in thread history, keeping the orchestrator’s context compact. NAC supports parallel execution through directed-acyclic-graph dependency management, and Arcee said the design takes inspiration from Random Labs’ “Slate” report.

Arcee published NAC’s code on GitHub, built by Brett Larsen, Lucas Atkins and the Arcee AI team. The company said it uses NAC internally for tasks such as running experiments, supervising training runs, working on infrastructure and prototyping ideas, and that NAC ships with an MCP server so a coding agent such as Claude Code or Codex can dispatch and steer its jobs.

Arcee’s account of its own internal use is self-reported and not independently verified, and NAC’s approach, which decomposes work and summarizes it into episodes, is one of several competing strategies for long-horizon agents, not a settled solution. The open license means other teams can now test whether the thread-and-episode pattern holds up outside Arcee’s own workflows.