Tmux + Fable = Cut 35% less token

[00:00:00]You probably experienced your cloud quota burning out Fable quota so fast that you hit a limit very quickly. And that's probably because you are not using the model right. So, as we get stronger and stronger frontier model like Fable 5 or GPT-5.6 so max, even though those model are extremely powerful, but also they are expensive and slow to run. That's why more and more people, even Cloud Code official, are recommending this practice that many use Fable 5 as the advisor or planner, but use smaller model like Sonnet 5 to be the executor. Cuz even though Sonnet 5 price is only 20% of Fable 5, but its performance is almost equivalent to Opus 4.8, which is strongest model a few months ago. And there are two ways Cloud Code official has shown. One is that using Sonnet 5 as the main executor agent, but only do it to call to get Fable 5 as the advisor to almost review its plan and send advice. Or second method is that you use Fable 5 as the main agent where it will make a plan and spin up worker agent using smaller model for execution. And even Devin is introducing this new harness type called Devin Fusion that claim to achieve similar or even better performance than Fable 5, but with 35% lower the cost.

[00:01:12]And fundamentally, their Fusion harness is just Cloud Code's Fable 5 orchestrator plus Sonnet 5 worker model.

[00:01:19]And Devin actually compare the two methods Cloud Code mentioned. One is the Sonnet 5 as the main executor, but use Fable 5 as advisor versus using Fable 5 as main orchestrator, but spin up a worker agent using smaller model. And the recommendation from them is actually very clear that the Fable 5 orchestrator model is working a lot better because with advisor model, you actually need to pay a very expensive price because advisor model will need to read the full conversation history from the main agent, rather than just trigger a cached context. And cached token is about only 10% cost of new input token. So, turn out using this model is actually going to work a lot better and cost-efficient.

[00:02:01]And to make it work well, the main ideas here is that each agent here should not be just a basic sub agent, but instead Devin is calling a sidekick. And what this means is traditionally, when you use Cloud Code or Code X and start a sub agent, it generally just spin up a new agent session on the fly. And once agent finish, the last message will be sent back to the main agent. But, let's say main agent realize there's something wrong with it. And when you want to make edits, the new task tool will basically spin up a new sub agent session that has no context of what has been discussed before, which means a lot of wasted token on rewriting the same thing. But, with sidekick model that Devin mentioned, the key difference here is a sub agent actually is a persistent session, which means if the main agent has some feedback, it will just send back a follow-up message to the same sub agent, which will inherit all the past context. And those past context, since we're being the cached token, it's going to be extremely cheap, yet very powerful, because it has a full context.

[00:03:01]And if you're using Cloud Code, you're probably already pretty familiar with this model, because that's literally what agent team means.

[00:03:12]You can literally just prompt Cloud Code to use agent team to do certain task.

[00:03:17]Instead of just starting a one-time sub agent, you can run this persistent session. Once it's finished, the main agent will be notified, and it will be continuously sent a follow-up message to this persistent agent session, instead of starting a new session. Cuz the new Cloud Code version already have the send message tool that allow you to send a new message to an existing agent session. And once those agent finish, it will stay resumable to receive those message until it calls a shutdown request. And when Cloud Code starting a agent team session, it can even define the model as well. So, So, you're using Cloud Code, it's actually very easy for you to achieve this type of Fable 5S main coordinator and planner and Sonnet 5 or even smaller model as executor. And you can achieve this workflow by simply adding this delegation rules into your cloud.md file. It's basically saying that you are the coordinator, that will be design, design, and review. And we delegate hands-on execution to the executor sub-agents using the Sonnet model. It lists out what are things that should be delegated and what are things that should be kept in the coordinator main agent. Which are specific things like design, planning, architecture, or just tiny edits. And there's one section that's pretty important about road check. So, the sub-agent would know it a sub-agent, so it don't keep spinning up more nested agents. So, with this one, I can simply start a session asking to help me design and build a to-do app. It will ask me question, do some planning, and typically for complex task, it will create a frozen spec in the task folder, then make the plan very clear, so there's no ambiguity for executor. Then spin up a sub-agent with Sonnet 5 to do the actual execution. And the sub-agent will receive this task saying the role is executor, do the work themselves, don't spawn sub-agents, and then read the spec first, and do the implementation. And because it is Sonnet 5, it's actually going to be much faster than you implementing with the Fable 5.

[00:05:10]And I have included my delegation rules in the AI Builder Club GitHub repo. So, you can copy and insert into your own cloud.md. If you're interested in learning more, there are more step-by-step tutorial on this coding agent workflow, and we have weekly live workshop to cover those. I have included link in the description below, so you can come and join if you're interested.

[00:05:28]However, this only works if you're using Cloud Code. But in reality, probably you have a few different AI coding agent that you're using. Some of you probably using Gemini CRI because the Gemini model is still one of the best and cost-effective model for front-end design, or the recent Kimi case 3 model.

[00:05:44]So, in an ideal world, you still want to use the best possible model and harness as a coordinator, which can be the Cloud Code or Codex using the GPT- 5.6 So Max.

[00:05:55]But the actual worker you should be able to free to use any of those based on your preference. And Codex team has been pretty sneaky. They introduced this Codex plugin CC tool, which is a Cloud Code plugin that allow you to delegate task to a Codex agent. So you just install this plugin and add the OpenAI Codex, then reload plugins. And now I can prompt it to delegate to Codex agent to help improve our UI of to-do app on this URL, which is first version that our Cloud Code agent did. And then it will start this Codex session. And here is a bit of weird cuz it basically start a sub-agent to call this Codex agent.

[00:06:33]It's probably because our coordinator/executor setup in cloud.md. But while it is running, you can see this Codex plugin giving us a few command. One is Codex rescue. This basically treat Codex as a sub-agent. Or a Codex review, which allow you to get a Codex to review the local git. And Codex result allow Cloud Code to fetch the output from a specific Codex session. Same as Codex cancel. And it can even do Codex transfer, which should basically take the conversation history to this point but fork over on Codex side. Now if I check my to-do app again, you can see that it update the UI, which looks probably more OpenAI-ish while the Entropic feeling. The one limitation with this Codex plugin is that I don't think there's a way for you to keep sending message to the same Codex session. If here I ask it to, "Can you send a follow-up message to this Codex session to say good job?" Oh, actually I am wrong. It actually can send back message cuz there's a resume command, which will allow it to send message to the existing session. So it can actually achieve that by doing this hack. So, the bridge between CodeX and Claude Code is pretty much there. But, what if you wanted to actually talk to any other coding agent that you are using like Gemini CLI, Guark Build, Pi Agent, Open Claw, any others. This is actually a way by using Tmux. So, if you don't know, Tmux is basically a terminal multiplexer. It allows you to start multiple different persist terminal sessions and continuously read and manipulate any other terminals. For example, I can run this Tmux split window -h, which means it will start a new terminal session on the right side of the screen. And inside that terminal session, I want to run Claude Arsay and with this message. It will spin up a new session on the right and send out this message. And what's cool is that on one hand, this session is persist. You can continuously sending more message to it.

[00:08:29]For example, I can do Tmux send keys t, which is the target, and dollar one means the second terminal sessions on the screen, "Tell me a joke." And then, I'll do Tmux send keys enter. This one will basically control the terminal on the right side to typing this message and click enter to send out message. And once I run that, you can see on the right side, a new message has been sent.

[00:08:54]And I can also run this Tmux capture pane -p -t.1. So, this means it will actually capture what's showing on that terminal and print out the result. So, if I click enter that, then it will read the result from the screen on the right side and return that result in the terminal. So, with this one, you probably start getting the idea that you can just tell your agent to use Tmux to start new agent session just like you as human. And through that way, it can technically control any coding agent and use that as a real agent teams. The only caveat here is that it's easy for you to start new session in this way. But, how can the the agent actually notify back the main agent that it has finished task or it needs some help. So, there are some hacks you can do. So, in tmux, there's a way for each terminal to send signals like by doing this wait for -s with a message. So, what do you technically can do is start a tmux session and instruct agent that once it's finished send this special message. And then run this tmux wait for this special message. You can even just pipe and printing out the section after it finish. So, if I do that, it will restart the new agent session and once it finish, it will run this tmux wait for down hello world page command so that this command will finish. And you can imagine you can basically instruct your agent to follow this pattern, run background command so that once it finish, it will be notified. And I actually created a skill called open agent teams. It's basically encapsulating this tmux hack alongside reference and adapter for each harness like cloud code, code ask, grok pi, open code and maybe others you can add to achieve those kind of agent team experience for any agent. And it actually kind of works in any terminal you're using. Like I can just open a normal terminal and start grok and just give a command. Ask code ask to make a joke about engineer who write code and after finish delegate to pi agent to review the joke and also cloud code high cool model to check the grammar in parallel. And after all finish, you do the final review and tell me. Right, so you can do things like this. And with this skill, I actually encapsulated the short script that is encapsulating all those tmux kind of practice but exposing like start a new agent session, send message to a session, wait for the result, see the result, things like that. So, you can see if firstly run this command, start a code ask session and this is kind of ID with a prompt, then it run this peak command to quickly check if this session actually running well Cuz sometimes those coding agent session can have those kind of permission confirmation screen. And in this case, it is working well, then it will run this command to wait for the session to finish with a max timeout.

[00:11:44]And the session is running in the background. And once finished, it will do this result check. And here you can see that the Codex actually wrote a summary, but the summary didn't include joke itself. So, then it can send a new message to a specific session to ask to print out the result. Then it got a joke, stopped the Codex session, start a pie review session for this specific joke. And meanwhile, also start a haiku grammar session with a special prompt.

[00:12:12]And you can see there's some issues that is blocked by the classifier. So, there's probably some improvements that we can do for the script as well, but it is able to just self-healing and wait for the result from both this pie review and haiku grammar check. And also see that the haiku grammar session is in this kind of interactive mode. Then it was able to just do the enter to control this terminal session. And in the end, return back the final version to me. And I have included this open agent team skill inside the AI Build Club repo here. So, you can just copy and use. It has this skill file talking about how to use this script and also a reference file for the cloud.md or agent.md file that including this delegation rule section. So, you can just copy paste over. But meanwhile, if you really want kind of premium experience, there are also two like Herd and Orca that has those orchestration CLI feature built in and pair with a kind of UI layer. So, it's pretty easy for you to see the result. And I test almost all those options. Personally, I found Orca is providing the best experience. So, Orca has this orchestration skills, which is a CLI plus Q come out of box. And you can basically install it. And once you install that, it is basically achieving the similar open agent delegations in there I have to here. But because it's very tied to the interface they have here, the experience became very intuitive. I'll give you one example.

[00:13:37]Let's say I just start a kind of new report, and in this report we don't have any uh skills. It just have that orchestration skill come out of box. I can basically tell the same thing, like help me ask Codex to make a joke about any new whoop code. After finish delegate to Pi, except I will also add the skill to just make it more explicit.

[00:13:57]Sometimes they will just do it by itself as well. But just for demo purpose, if I enter this, it will use the skill which will invoke Orcas kind of still I that they build in. And then first they start a Codex session. But the really cool thing here is that you can see this new session just pop up on the right side.

[00:14:16]So, it's actually very easy for me to see all the sessions end-to-end, and I can just put the sub-agent session on the same screen here. Uh and on the left side, I just have this kind of hierarchy view to show me what are the sub-sessions under the session. The same thing, it also start this Pi agent for the review, as well as this Haiku agent for the grammar check. And again, everything is just very well tracked in their interface. I can just go into any sub-agent to continue the conversation there, and uh once I finish it will like automatically communicate back to the main agent. And I believe they will even have some kind of special data visualization to help you see how the tasks orchestrated across different agents. So, if you want a more kind of packaged experience, I highly recommend you check out Orcas well. They do have so many good stuff built in, like the Orca mobile experience, the Kanban view out of box, the token usage tracking, all the good stuff. And it's totally open source and free, and they're growing like really, really fast. For the past few weeks, it has been my daily driver for communicating with agents for the past few weeks. And we actually did a 1-hour deep dive workshop on some of advanced AI coding workflow with Alka founder. We actually showcase step-by-step of how they were managing to ship hundreds of commits to the Alka product itself. In Octrade advanced agent workflow across Grok, Codex, Cloud Code, Open Code. You can also join AI Build Club to see the in-depth workshop.

[00:15:42]So, I hope you enjoyed this video. As mentioned, feel free to get the skills that I put in for either using this open agent team skill as well as this Contagion Rules section for your cloud.md or agent.md file to achieve those kind of Octrader worker setup. So, that's pretty much for it today. Thank you and I see you next time.