Today we're open-sourcing nac, an agent harness for long-running tasks, and launching the Arcee open models API beta. Nac is available now under Apache 2.0 on GitHub, built for developers running complex, multi-step engineering workloads.

A central orchestrator plans work and dispatches it to isolated worker threads. Workers execute in clean contexts, make their edits, and return a concise summary—an episode. The worker process exits and its raw context is discarded, while the episode persists.

Threads can weave episodes between each other to share context without carrying historical noise. Nac also ships with an MCP server. Interactive coding agents like Claude Code or Codex can use nac as a tool, delegating long-running background execution while keeping your interactive chat fast.

To support these workloads, we are expanding our API beyond Trinity with the Arcee open models API beta. Long-horizon tasks require different model strengths at different steps. Hosting a broader catalog gives you choice, while helping us learn how frontier open models perform in persistent agent runtimes.

Launch catalog & pricing per 1M tokens (input/output): • deepseek-v4-flash-latest ($0.14 / $0.28) • trinity-large-thinking ($0.25 / $0.80) • thinkingmachines/inkling-small ($0.50 / $1.20) • deepseek-v4-pro ($1.74 / $3.48) • zai-org/glm-5.2 ($1.40 / $4.40) • moonshotai/kimi-k3 ($3.00 / $15.00)

Try it today with $5 in credits. Sign up for the Arcee API and add a payment method to get $5 in API credits to test nac against the new catalog. • Platform: • Open models beta blog: • Nac repo: • Nac blog:

Today, this is a mix of internally hosted endpoints and external partners like @FireworksAI_HQ for Kimi-K3 (via their agreement with Kimi), allowing us to serve beyond our current internal capacity. We do not monitor or train on API outputs, nor do our partners as a part of the agreement.

Credit to our team for shipping both today, and shoutout to @deepseek_ai, @Zai_org, @Kimi_Moonshot, and @thinkymachines for building great open models. We build harness runtimes and open model infrastructure so you can own your AI stack.

They said they wanted options and you gave them options. Amazing job, Arcee team. We're biased towards K3, of course, but you know. Try 'em all.

lfgggggg

This looks great, the team should publish nac on the @swarms_corp marketplace to gain more potential users and awareness

This is just what I was thinking about today!! You guys just saved me a bunch of time from building it myself. Thank you!

wait it works with codex auth too?

long-running agents finally got a seatbelt and a flight recorder. how are you checkpointing tool state across model swaps?

i love an open-source agent harness that does great for long-horizon complex tasks! congrats and well done team! 🤍

Holy launch video

Nac as Apache 2.0 harness for long-running tasks lands squarely where the infrastructure opportunity just shifted. The identity layer and agent harnesses are now the hot sector, not training. I'd watch whether Arcee's harness handles attestation across recursive loops, or if that's punted to the caller.

kino

The beta of that API still has to talk to the tools its supposed to augment. What does the integration spec look like?

Release the rest of it, cowards. Release ogdoches.

Long-running agents are useful right up to the point where they overwrite something you cannot audit. Elements in EvaEsi reads first, proposes the change, then waits for your approval. Launching later this year.

@grok 这个纯研究还是有工业意义，具体工业场景视角看意义是什么，有开源数据集或者开源项目代码吗？从多个数据源交叉验证，不要只看新闻媒体一面之辞。帮我排除没意义的垃圾商业营销推广、诈骗、夸张博眼球、虚假新闻 以及自吹自擂，自嗨，无病呻吟。

Looking forward to seeing @superiorsi competitors post-train on @arcee_ai bases!

开源长期任务代理框架对开发者很友好，落地门槛会明显降低

nac will be judged by checkpoint recovery because long-running tasks rarely fail cleanly

An agent harness for long-running tasks lives or dies on durability, not the loop. The real test isn't reasoning - it's resuming a 6-hour run from a checkpoint after a crash instead of restarting from zero. #AgenticAI @arcee_ai

Isolated workers help until the episode is a lie. If the worker exits after a half-write, the orchestrator just schedules the next half-write.

nac looks like a solid foundation for long-running agent workflows. Apache 2.0 means teams can embed it without license friction.

Long-running agents need more than a loop: durable checkpoints, explicit task state, and artifact-level traces. Otherwise a 2-hour job becomes an opaque chat session with extra failure modes. Making the harness open source is a meaningful step toward resumable, inspectable work.

long-running agents fail at recovery, not planning; does nac persist tool state across restarts

Why everyone is releasing harness today

Long-running agent tasks are where this kind of tooling gets really interesting. Open-sourcing the harness is a strong move for developers. 👀🔥

Let's connect for for collaboration🚀

Apache licensed agent harness for complex workflows is huge. We actually covered the same momentum with DeepSeek here:

Apache 2.0 is always a good sign