Some of those points resonate with what we're experiencing with our current approach of running agents and sub-agents locally.
- We are delegating complex and long running tasks to agents more and more. Being able to run those overnight seems like a logical next step in efficiency, because at the moment we are stopping the local agents every evening and restarting them in the morning.
- A setup with decentralized multi-agent system with shared context is 5.5x faster than Claude code + native subagents. It takes 29.3 minutes (2% of agent time waiting) versus 162.7 minutes (44% of agent time waiting). [3] This aligns with our feeling that locally running agents are often "doing nothing" while waiting on test results, each other, and so on
- The possibility to monitor "input channels" (e.g. chat channels, e-mail inboxes...) and have agents act autonomously on those seems like something that could help us out as well.
This all sounds amazing, but we're unsure how others are currently doing this and what those cost. We currently mainly power our needs via subscription-based contracts (e.g. claude (code) pro/max, chatgpt plus/pro ...)
- Is anyone running such multi-agent or always-on agent systems?
- Do you host it yourself, or do you use existings solutions like grok bots or claude remote control sessions or claude routines?
- Can you make this work with subscriptions, or do you use API pricing, or even self-hosted LLMs? How do the costs compare with running local agents?
- How are the results? Is it an improvement?
[1] Stanford CS146S: How always-on agents work
Lecture: https://www.youtube.com/watch?v=L1OBRtfPS34
Slides: https://drive.google.com/file/d/1b8WiBk_lcB3ue-jZMLijSUMLj6SDiF6Z/view
[2] CMU 11-768 AI Agents: https://www.cmu-agents.com/#/schedule
[3] Decentralized Multi-Agent Systems with Shared Context (figure 3, page 3): https://arxiv.org/pdf/2606.10662
[–] naw103 3 hours ago
[–] abejora 1 day ago
[1] Stanford CS146S: How always-on agents work
Lecture: https://www.youtube.com/watch?v=L1OBRtfPS34
Slides: https://drive.google.com/file/d/1b8WiBk_lcB3ue-jZMLijSUMLj6S...
[2] CMU 11-768 AI Agents: https://www.cmu-agents.com/#/schedule
[3] Decentralized Multi-Agent Systems with Shared Context (figure 3, page 3): https://arxiv.org/pdf/2606.10662
[–] abejora 1 day ago
[–] morisblake 7 hours ago
[–] harlequinetcie 1 day ago
Particularly interesting for you, our lasts posts on how we handle to have efficient dreaming focused on reducing costs.
tl;dr: there's a self-driven process generating more effective and efficient prompts, code, and potentially select systemone models to tackle the cost problems.
We have companies using the system that we understand have stayed 'flat' in token consumption but with increasing workloads (ie. more workflows running).
Hope it helps!
The part that took the work was shared state. We used GPTree for the knowledge engine. Every routine reads and writes from the same knowledge and anything public or high risk gets an approval card to continue.