Pi Agent vs OpenCode token usage A lot of people recommended Pi Agent so I decided to check Pi Agent took 1.1k tokens in first turn OpenCode took 11.5k Setup: 1) Trimmed OpenCode (from usual 30k first turn to 11.5k) - 0 MCPs - 2 lightweight plugins (opencode-env-protect and openslimedit) - 8k char AGENTS.md - 11585 deepseek-v4-flash input tokens - for $0.0016 2) Vanilla Pi - 0 MCPs - 0 system prompt - 8k char AGENTS.md - 1114 kimi-k2.6 input tokens - for $0.0008 I think using better models with capped tokens per turn can keep usage nearly the same as uncontrolled DeepSeek V4 Flash? The challenge now is finding the sweet spot. Can't cap tokens if quality drops. We'll see. Video below - OpenCode vs Pi side by side "say hi back" first turns test, with OpenCode Go usage for each

For a weekend hobby experiment, as a tinkerer I would say yes But launching it as product? Let's be real, only if the incentives make sense A new venture? A client paying for R&D? A partnership? A couple of warm buyers waiting to try for their teams? A profit sharing plan? Cause I would think that it's at least one of those things What separates a fun weekend project from the real thing is the mindset shift from the fun vibe of tinkering to operating a product that users rely on With great power comes great responsibility Ship more? Cool. Expect to review more, support more, and fix more. Could be you and your fleet of AI agents, or you and your team of cracked humans with their agent fleets working 24/7 piling up inference bills. But all that it takes is a massive defect from a drastically incorrect implementation to strain your dev pipeline So at least be prepared to handle those things

The new OpenCode dashboard looks promising but the problems are real My 30 day log history from the old dashboard (legacy) is still gone. They assumed fresh use and just logged new requests? The more concerning one: API keys are failing. Both legacy and dashboard v2 api keys are failing on inference attempts. Even with an updated v1 client. Inference only works if I use their new auth system in 'opencode console login' which logs you as a user not an api user This roll out should be opt-in I also hope they don't force everyone to upgrade to opencode v2 either Unless it's stable enough and the migration path is smooth as butter

As a MiMo V2.5 enjoyer I'm very excited for MiMo V2.6 I consider MiMo as that balanced, smooth as silk model because it plays nicely with most of my workflow: From system prompt, skills, up to the way I prompt and converse with my agents MiMo is just one of those models that does not fight with my setup, a cheaper model that converses with the simplicity of Claude and the structure of ChatGPT GLM and DeepSeek are great, but GLM is too verbose and technical to my liking, which is essential don't get me wrong, but not great when you just want direct to the point responses and a path towards clarity, not more complexity. On the other hand, DeepSeek loves adding "one more thing" like Steve Jobs and still misses stuff despite doing that, due to its notoriety of scoring so high in AA-Omniscience Hallucination Rate. Don't get me wrong, that makes DeepSeek a creative model, at least from my own experience conversing with it. However, MiMo is known for having low hallucination rates just like Gemini 3.8, Kimi K3, and GLM 5.3, and these are even lower than Claude models which frequently sits at 60% range. But what makes me most excited is MiMo V2.6's price and speed I/O token price comparisons show: ~1.5x/2.4x cheaper than GLM 5.3 discounted prices ~3.4x/8.6x cheaper than Kimi K3 Artificial Analysis charts show 2x speed I am recently impressed with DeepSeek APIs speed. If Xiaomi MiMo keeps up with being half as fast, it's a no brainer that I will use it more. Faster inference just keeps me in the flow state longer, and once you experience that, it's harder to tolerate slower models.

Probably one of the biggest constraints in AI integrations is the integrations themselves. I need to work on a set of hundreds of Google Drive hosted videos, by moving them somewhere else that my app can play reliably. First, I need to get their start and end times as they were embedded on the video container inside a Google Slide. Claude is the best at parsing slides but it cannot natively access Google Slides video containers nor manipulate it with Apps Script. However, Gemini can do this easily. So I have to detour with Gemini and Apps Script. Problem solved. Second problem, I have to get their links so I can batch download them. Back to Claude, who in turn gave me a clean markdown of links ready for curling. Problem solved. Third, I have to set up a script to upload them on the bucket that my app is using, and plug the links to my app's database, along with start and end times fetched earlier from Google Slides' video containers. This is the final step. So I handed off the Claude markdown to DeepSeek that's powering my local harness. That's the one building my app. It has access to my local workspace that both Claude and Gemini web chats can only dream of. At this point I already used three AI chats, with their own agents and tools underneath, with native integrations that's basically their moat. I wonder how a multiplayer agentic workspace can change this. Imagine not having to juggle between chat apps. Better yet, imagine not having to orchestrate these cause the agents simply knew what to do.

Late to the party but I hope I'm not late to the waitlist In the meantime I'm going to check the open source alternatives popping off atm Lots of use cases

off peak deepseek hours and 150 tps is just addicting almost burned half of my $3 credits in just 2 days

One rule that I always do with testing providers I'm unfamiliar with: Take it as an experiment. Small credits first, and test it for a while on low bearing projects. Still painful to see that I spent $4 on this but the inference speed will blind you. These things can happen to anyone even with established providers, so hesitation and common sense comes a long way. The UI is the clue though. It's so unusable that it reminds me of Chutes infra. I'm lenient to these things, as teams break things fast but fixes things fast, or maybe just a solo guy running it bts, but the messaging that a very unpolished product gives off will always invite hesitation, believe it or not. At least Crof refunded my remaining balance.

Multi agent apps are hot again thanks to Grok Bot. So I thought of having one locally for fun. Who doesn't love having the same user experience, with control on the provider you use? A good starter is taking local sandboxing seriously, which I lack as I rely on permission gates more and doing most of my work on my local machine. But interacting with your agent 70% of the task is not it. The real unlock is the time and complexity you freed, isn't it? So I'm taking "hands free" seriously this time, and it starts with containment. Steering comes next, but that's for another day. If you're on Windows, containment starts with a minimal WSL 2 instance and Pi agent. If you love NodeJS you better have nvm with you and if you love Python, it has to be uv. Now, here's a shortcut: You can skip Podman and Docker if you start your WSL 2 sandbox with only your project folder mounted to it. You can easily do that with your main agent harness controlling your host machine, or a script that injects into /etc/wsl.conf inside that WSL 2 sandbox. Could be OpenCode on Orca if you love the CLI look, or Paseo if you like the desktop feel. Or just a simple .cmd/ps1 launcher script. Up to you, if you like Batch or PowerShell more. Of course, you have to launch your WSL 2 sandbox with a non root user so you're still "sudo" protected, plus I/O piping for Pi agent, and "exec pi" in the end. Sudo might not be enough, but Linux is well equipped with user permissions that worth exploring in another day. That's it. I just finished the first step towards my local Grok Bot goal. The first test is the image below: An OpenDesign instance, using a Pi agent inside my WSL 2 sandbox. No more permission gates. No more host altering changes. Everything is contained in a WSL 2 sandbox and just the project folder itself. So, if OpenDesign can use a harness contained in a sandbox, which eliminates the need to ask me to intervene, then any harness can, like an open source Grok Bot, which I am going to explore next. I will definitely see how Podman or Docker fits in the whole picture, as I want more predictable, recurring workflows mimicking containerized production environments. All it took is half a day of setup and with the help of DeepSeek V4.1 Flash during off peak hours. On a weekend of course. Tweaking is fun.

Probably one of the most underrated aspects of realizing value is taking that extra step. Yes it definitely looks like a distraction, but that could also be the next failure you need. Your personal Kanban mission board is almost a multiplayer agent playground. An old school test suite is almost an interactive testing playground that you can ship out to human and agent testers. Your bunch of local machine permission gates is almost a permission-less "hands free" sandbox with multiplayer agents using it. Almost. Almost. That's in your hands now. Sure, definitely takes more effort and costs more $$ with Astra, Fable and a bunch of SaaS and API connectors, or probably just a lean VPS, and it definitely eats more of your time if you go with cheaper, less capable models. Of course it will fall in obscurity if you don't share it to your friends, clients, colleagues, and your neighbors. Let it loose and ask people. At the end of the day, whether you do that or not, what do you learn with that extra step? Or at least the fact that you took that extra step? It's addicting once you start, and it's definitely invaluable. Makes you go places isn't it?