0:02 Hi, my name is Mario. I hail from the 0:04 land of Arnold Schwarzenegger. Uh, which 0:07 you probably haven't noticed yet based 0:09 on my very good English. Uh, I want to 0:12 preface this with we've been running 0:14 around with our four-year-old the entire 0:16 day through London. So, we went to 0:18 dinosaurs mummies Nandos obviously 0:21 uh, 0:23 uh, and stuff I have already forgotten. 0:26 I'm very, very tired. And if uh you 0:29 don't understand anything I say uh just 0:31 raise your hand and and say grandpa wake 0:34 up you um the reason I'm here is 0:38 actually another person which is here in 0:40 cognto today let's call him uh 0:43 Pineberger 0:48 um back in 2025 I think somewhere around 0:51 April 0:53 he told me and Armen Rona which you 0:57 might also know Flask fame and Sentry 0:59 fame. Dude, those coding agents, they 1:02 actually work now. And I was like, oh, 1:05 shut the [ __ ] up. Sorry, I'm also using 1:07 swear words. Um, totally not. Uh, and a 1:10 month later, we teamed up at this flat 1:12 for 24 hours overnight and just let 1:15 ourselves uh get um immersed by the 1:18 clankers, by the wipe coat, and by the 1:20 wipe slop. And since then, none of us 1:22 have really were sleeping anymore, 1:25 basically. So we're building stuff, lots 1:30 of stuff, most of which we actually 1:32 never used because that's the new thing 1:34 in 2012526. 1:37 We build a lot of stuff, but we don't 1:38 build a lot of stuff we actually use. We 1:40 wrote a lot of stuff. Uh and eventually 1:44 that culminated in me thinking, I hate 1:47 all the existing coding agents or 1:49 harnesses. How hard can it be to write 1:52 one myself? And Peter was like, "Oh, I 1:55 just want to do a thing. Nobody's 1:57 probably going to hear about it." And uh 1:59 it's going to be a personal assistant 2:01 because that's what I always wanted to 2:02 have. And yeah, most of you probably 2:04 know how his story went. So today I'm 2:06 going to tell you my much less 2:07 impressive story, but I I hope I can 2:09 transport a couple of learnings as we 2:12 say in the industry um that I was able 2:15 to gather in the past couple of months. 2:18 So pi in the beginning there was cloud 2:20 code actually there was copy and pasting 2:23 from chatpt right we all did that in the 2:26 beginning 2023 uh then there was who 2:29 remembers the original github copilot 2:32 yeah actually how many of you are 2:33 engineers 2:35 how many of you are using coding agents 2:37 like cursor cloud code okay uh 2:40 popularity contest cloud codec 2:44 cli 2:46 cursor reversals. 2:49 >> Yeah, open code 2:50 >> anti-gravity. 2:52 >> Oh, that's not a lot. Anybody using 2:53 this? 2:55 >> I like you. We're going to have a beer 2:57 later. Anyway, so this was basically uh 3:01 what happened right in 2025 and before 3:04 start with copy and pasting from CHPd. 3:06 It's all mostly broken. It's mostly 3:07 single functions stuff you don't want to 3:09 write. Then you got GitHub copilot 3:11 inside of your Visual Studio Code where 3:13 you just tap tap tap to happiness which 3:16 did work sometimes mostly didn't. 3:18 Sometimes it will also just patimly um 3:21 recite GPL code like John Carmeck's 3:24 inverse square root and stuff like that 3:26 which was a lot of fun. And then there 3:29 was Ader. Anybody remember Ader? Yes, 3:32 old people. Hello. 3:35 Um yeah, you have gray hair. You 3:38 obviously know Ader. Um there was also 3:41 autog. 3:44 Yeah. Okay. He knows all the things. So 3:47 um and then eventually there was cloud 3:50 code. Um I think they released it in 3:52 November actually as a beta in 2024 but 3:55 it really only became used uh more 3:59 >> early February. 4:00 >> Yeah. February, March, something like 4:02 that. 2025. And I was like I love it. 4:06 It's awesome. The cloud team is also 4:07 awesome. on their own socials and 4:09 they're all very good people and very 4:11 talented people. Um, and they basically 4:14 created the entire genre. I know there 4:16 were precursors like Ader and AutoGBT, 4:18 but nothing did this. And this was 4:20 basically the whole aentic search thing. 4:22 So instead of like cursor going into 4:25 your codebase, indexing things, 4:26 constructing a indexing that as well, 4:29 and it's kind of not really working. Um 4:31 they just said we reinforcement trained 4:34 our models to just use file tools, bash 4:37 tools to explore your codebase ad hoc 4:39 and find the places that it needs to 4:41 find to understand the code and then 4:43 modify the code. And this worked so well 4:46 that yeah we stopped sleeping because we 4:48 all of a sudden could produce so much 4:49 more code than we could before by hand. 4:52 Back then it was simple and predictable 4:54 and uh actually fit my workflow 4:56 perfectly. um fine but then 5:01 they fell into the trap and to which 5:04 most of us us probably fall. The 5:06 clankers can write so much code why not 5:08 just let it write all the features you 5:10 could ever imagine, right? Isn't that 5:12 great? Let's just add this feature and 5:14 that feature and this feature and that 5:15 feature and eventually you end up with 5:17 uh Homer Simpsons I don't even know what 5:19 it's called. I call it a spaceship and 5:22 cloud code is now a spaceship. It does 5:23 so many things that you actually 5:25 probably ever used like 5% of what it 5:27 offers. You only know about 10% in 5:30 total. And the rest, the 90% that's left 5:33 over, that's kind of like the dark 5:34 matter of AI and agents. Nobody knows 5:36 what it's actually doing. And I 5:38 personally find this not to be very 5:40 helpful because I still think that you 5:42 kind of need to know what the agent is 5:44 doing. This guy might disagree to some 5:47 degree. Um, and we're here at Tessle and 5:51 they also like context management, 5:53 right? Or context engineering as we've 5:55 called it. And I eventually found that 5:57 cloud code was not a good tool when it 6:00 comes to observability and actually 6:01 managing your context. 6:04 Um, then there was also this. Who likes 6:06 this about cloud code? Like the immense 6:08 amounts of flicker, unexplainable 6:10 flicker. Well, actually I know how to 6:12 explain it and why it happens, but they 6:13 still haven't fixed it. Here's TK. He's 6:16 really great. I I love him. He is their 6:18 defrail guy mostly on Twitter and he's 6:20 amazing. But but sometimes he also says 6:23 questionable stuff like our terminal 6:25 user interface is now a game engine. Now 6:28 you have to know I have a game 6:29 development background. Like that's 6:31 where I come from. And if I read 6:32 something like this then it kind of 6:33 hurts me a little bit because it's a 6:35 freaking terminal user interface. Dude, 6:37 it's not a game engine. Trust me. The 6:40 only reason you think it's a game engine 6:41 is because you're using React in your 6:43 terminal interface and it takes like 12 6:45 milliseconds to relay out your entire 6:47 user interface graph. Just don't do 6:50 that, man. You don't it's not a game 6:52 engine, right? So, and then uh Mitchell 6:55 who is writing Ghosty was like, "Dude, 6:58 that's offensive, man. Like, don't blame 7:00 it on Ghosty or any other terminal. Your 7:02 code is garbage. 7:04 terminals can render at like hundreds of 7:06 frames per second uh sub milliseconds 7:08 per frame. So don't do that, right? Um 7:12 and then they eventually fixed the 7:13 flicker, but then other stuff happened. 7:16 So it's like they fully gave into the 7:19 vibe coding and you can feel it every 7:22 day when you use cloud code. Now again, 7:24 I do not want to diminish their their 7:26 efforts and their results. Claude Code 7:28 is still the category leader for a good 7:30 reason. They invented this thing and 7:31 they're doing a great job. I personally 7:33 I'm just an old person who likes 7:35 predictable simple tools uh and this 7:37 just didn't fit my workflows and my 7:39 needs anymore. So yeah, also they do a 7:43 lot of stuff in the background 7:45 manipulating your your context. I built 7:47 a bunch of tools in summer 2025 that 7:50 would allow me to intercept requests 7:52 being made to their back end from cloud 7:54 code and finding out what kind of little 7:57 additional text gets injected into your 7:59 context behind your back and all of that 8:00 was very detrimental and it also changed 8:03 all the time like every day every day or 8:05 second day there would be a new release 8:06 where this changed what gets injected at 8:09 what point which would basically mess 8:11 with your existing workflows. it was 8:13 just not a stable tool. And now I 8:14 understand it from their perspective. 8:16 They need to experiment and they have a 8:17 huge user base and it's really hard to 8:19 experiment when you have a huge user 8:20 base. But they did not care. So all of 8:23 us had to suffer, right? You you're 8:26 working with this new tool. You try to 8:28 create predictable workflows and then 8:31 the tool vendor changes a tiny little 8:34 thing under the hood that makes the LLM 8:36 go crazy with your existing workflows. 8:38 That's just not sustainable. I need 8:40 control over that. I can't rely on them 8:42 providing me a stable kind of thing. Um, 8:46 so 8:48 I believe as a consequence of the UI 8:50 design, um, they need to reduce the 8:53 amount of visibility you have. I 8:54 personally don't like that too much, but 8:56 that's just a personal preference. I 8:57 understand that most people will be 8:58 happy with the amount of information 9:00 that cloud code will present you. Um, 9:03 there is zero model choice obviously 9:06 because it's a anthropic native tool so 9:09 to speak. That's not a downside because 9:10 cloud models models are I like them like 9:13 they're really good. Um, and there's 9:16 almost zero angstensibility and you 9:18 might find this kind of funny because 9:19 they have this whole hook system and all 9:21 of that, but if you compare it to what 9:23 PI allows you to do, it's it's not as 9:26 deeply integrated. Um it's also 9:28 basically based on running a process 9:31 when the hook event uh starts which is 9:34 very expensive um if you have to start 9:37 up that process over and over again. So 9:40 eventually I I I soured on cloud code 9:42 not because it was terrible it's just it 9:45 did it stopped being a fit for me. It 9:47 became a fit for a lot more people over 9:50 that period. So obviously they they're 9:51 doing things right right but not for me 9:54 because I'm old. 9:56 So then I was looking around for options 9:59 and there's codec cli which I really 10:01 didn't like in the beginning both the 10:03 user interface as well as the model that 10:05 has changed at least with respect to the 10:07 model. Codex is really pretty good now. 10:10 Then there's AMP the team behind that uh 10:13 used to work at source graph. They 10:16 how do you they spun off of source graph 10:19 and they're super good engineers. they 10:21 they they managed to build a commercial 10:24 coding harness where they take away 10:26 features instead of adding them and most 10:29 of the choices make a lot of sense to 10:32 me. Um so yeah, if you're looking for a 10:34 commercial coding harness, I would 10:36 definitely recommend AMP to you because 10:38 it's really good. Factory trod kind of 10:41 similar spiel um also really good 10:44 although they're not as experimental as 10:46 AMP. And then there's open code which is 10:49 the open source uh coding harness a lot 10:52 of people use right so I have a history 10:54 with open source I've been in open 10:56 source for well uh 17 years I've managed 11:01 big and small open source projects uh so 11:04 that's near and dear to my heart and so 11:06 I thought I give open code a try because 11:09 that's close to me right 11:12 and next to AMP they have one of the 11:14 most grounded or pragmatic teams in the 11:16 space they don't hype you up with 11:18 features you probably never use. They 11:20 try to um kind of conserve a happy path 11:23 that's very stable. Um and they also 11:26 have pretty good thoughts on what coding 11:28 agents mean for us as a profession which 11:30 I personally can identify with. Um 11:34 the problem with open code is that it's 11:36 also not very good at managing your 11:37 context. Uh for example, on each turn 11:42 it's calling session compaction. Prune 11:44 which does the following. It prunes all 11:47 two results um before the last 40,000 11:51 tokens. Now, who here knows what prompt 11:54 caching is, 11:56 >> right? What does this do to your prompt 11:58 cache? 12:01 >> Yes. So, Open Code and Entropic had an 12:04 interesting history and eventually 12:07 Entropic in my opinion um rightly so 12:11 said, "Dudes, that's just not going to 12:13 happen." Right? And there was never a 12:16 public kind of thing about this, but Tar 12:18 explains it here. If you come to a gym 12:20 and don't behave and and abuse the 12:22 infrastructure, so to speak, you're 12:24 going to get banned. And I think I don't 12:27 have any evidence for that, but I think 12:29 that's the reason why there is this 12:31 animosity between Enthropic and Open 12:33 Code. And I can totally agree or at 12:36 least I think that Anthropic is clearly 12:38 in the right here. Um, don't mess with 12:40 the infrastructure. 12:42 And then there's also other stuff like 12:44 open code comes with LSP language server 12:46 protocol support out of the box. Coming 12:48 back to context engineering. Let's say 12:52 you give your agent the task of 12:54 modifying a bunch of files. What does 12:56 that mean in practice? It will make a 12:58 bunch of edits um one of the the other 13:02 to a bunch of files. How probable is it 13:06 that after the first edit out of 10 13:08 edits, so to speak, the code will 13:10 compile? What happens if you modify your 13:13 code line by line? How long does it take 13:16 for it to stabilize again and it 13:17 compiles cleanly? It doesn't. It won't 13:20 compile after the first edit, probably 13:21 not after the second edit, and so on and 13:23 so forth. So, if you then turn around 13:25 and say, "Hey, dear LSP server, I just 13:28 edited one line in this file. Is it 13:30 broken?" Then the LSP ser yes yes it's 13:33 really bo camp and what this feature 13:35 does it it then injects this error 13:37 directly after the tool call as a kind 13:41 of feedback to the model oh what you 13:43 just did is wrong and the model is like 13:46 what the [ __ ] dude I'm I'm not done 13:48 editing things why are you telling me 13:50 this obviously it's not wrong but if you 13:51 do this often enough the model will just 13:53 give up and that leads to very bad 13:55 outcomes um so I'm not a fan of LSP I 13:59 think it's a very terrible idea idea to 14:01 have that enabled. There's natural 14:02 synchronization points where you want to 14:04 have linting and type checking and all 14:06 of that and that is when the agent think 14:08 it's done only then. 14:13 This has changed recently. This is a 14:15 single session of open code um 14:19 every message becomes its own JSON file. 14:22 Every single message becomes its own 14:24 JSON file on disk. 14:26 That indicates to me that there wasn't a 14:28 lot of thought put into the architecture 14:30 of the whole thing. And if I lose trust 14:32 in that, I don't want to use that tool 14:34 anymore. Again, I think the team is 14:36 actually really good. I think they 14:38 iterated super quickly and built 14:40 something that's super useful to a lot 14:42 of people. Obviously, it's just again 14:44 decisions that I wouldn't have made that 14:46 made me decide to build my own. Then 14:50 there was also this open code comes with 14:52 a um a server by default. So the core 14:55 architecture is based on a server and 14:56 clients connect to it and the terminal 14:58 user interface is one of the clients. 15:00 There's also a desktop interface and I 15:02 don't know that turned out to be a 15:04 security vulnerability with remote code 15:07 execution baked in by default and that's 15:09 also like if you're so proud of your 15:12 server infrastructure or server 15:13 architecture then I would assume you're 15:16 grownup engineers that thought about 15:19 security as well. And apparently that 15:20 didn't happen. And this was open for a 15:22 long time. And again, I'm not blaming 15:25 anyone here. This is stuff that just 15:26 happens if you're working in an industry 15:28 that's operating at a pre break neck 15:31 speed that we haven't seen before, 15:32 right? It's just I don't want to use 15:34 that tool if that is a thing. 15:38 So, so this was my observations with 15:40 regards to existing coding horses. AMP 15:42 Android android would have been 15:44 something I could have used but again no 15:47 control in case of AMP they even decide 15:49 what models you can use and it's only a 15:51 single model for a single type of task 15:53 and that's not me in terms of I think 15:56 it's a little bit more open but at the 15:58 time when I tried it out um it just 16:00 didn't I I didn't see a big um advantage 16:04 over cloud code 16:07 and then I looked into benchmarks for 16:09 entirely different reasons and found 16:11 terminal bench Who knows what terminal 16:13 bench is? Okay, basically it's a coding 16:16 or an agent 16:19 evaluation harness which uh has a bunch 16:22 of computer use and programming related 16:25 um um 16:27 sorry old and tired because we're old. 16:31 It has a bunch of computer use and uh 16:33 coding related tasks that an agent or 16:36 the LLM inside an agent harness um needs 16:39 to uh fulfill. I think it's about 20 uh 16:42 82 or so and they're very diverse. 16:44 They're from fix my windows setup to um 16:47 code me a Monte Carlo simulation or 16:49 something like that. Um and they have a 16:52 leaderboard and on that leaderboard you 16:54 see the combination of coding agent 16:55 harness and model right and they have 16:59 their own coding agent called Terminus. 17:03 And I think it's brilliant because 17:06 it's one of the best performing 17:07 harnesses in the benchmark. We're going 17:09 to see it later on. Uh what exactly does 17:12 it do? Well, all the model gets uh is a 17:16 T-max session and all it can do is send 17:18 keystrokes to it and read back the the 17:21 VT code sequences that are emitted. So 17:23 this is like the smallest most minimal 17:26 interface a model can have to to your 17:30 computer, right? And this performs like 17:33 top of the line of the entire 17:34 leaderboard. 17:36 So what does this tell us about existing 17:38 coding agent harnesses? Do we need all 17:40 these features for the models to 17:42 actually perform? 17:45 For me personally, this is not just 17:47 about the model actually being good. 17:49 It's also about me as the user, the 17:51 human having a way to interact with my 17:52 agent with the model. And Terminus is 17:54 obviously not the user experience or 17:58 developer experience that I want. But it 18:00 tells us that all of these features, all 18:02 of these coding harnesses have might not 18:04 nec might not be necessary to um get 18:08 good results out of agents. So no file 18:10 tools, no sub aents, no web search, no 18:12 nothing. Two thesises based on all of 18:15 these findings. We are in the messing 18:17 around and finding out stage and nobody 18:19 has any idea what the perfect coding 18:20 agent should look like or what the 18:22 perfect coding harness should look like. 18:23 We're trying both minimalism and going 18:25 full spaceship swarms and teams of 18:28 agents and no control and full autonomy 18:30 and whatever. I think that's not done 18:33 yet. We haven't answered the question 18:35 what this should look like ideally and 18:36 what will become the industry standard. 18:38 And the second thing is we need better 18:40 ways to mess around uh with coding 18:42 agents. That is we need them to be able 18:44 to selfmodify them uhelves and become 18:47 malleable. So we can quickly experiment 18:49 with ideas and see if this is something 18:51 we can make like an industry standard a 18:53 new workflow that we probably all going 18:56 to adapt. 18:58 So the basic idea was and it's very 18:59 simple and not rocket science. Strip 19:01 away everything and build a minimal 19:03 extensible core 19:05 with some creature of comforts. It's not 19:07 a blank slate. So that's pi. Um and the 19:11 general motto is uh adapt your coding 19:13 agent to your needs instead of the other 19:14 way around. 19:16 It comes with four packages. U an AI 19:20 package which is basically just a simple 19:21 abstraction over multiple providers 19:24 which all speak different transport 19:26 protocols. Um so it's very easy to talk 19:28 to all the providers and switch between 19:30 them in the in the same context or same 19:32 session. Um the Asian core which is just 19:35 a generalized agent loop with tooling 19:37 locations verification and so on and so 19:39 forth. Then streaming um uh a terminal 19:42 user interface that's like 600 lines of 19:44 code 19:46 and works really well surprisingly uh 19:48 because it wasn't written by a clanker 19:50 um and the coding agent itself which is 19:53 both an SDK that that you can use head 19:55 in headless mode um or uh a full 19:59 terminal user interface coding agent. 20:02 This is the entire system prompt. 20:05 There's nothing more there compared to 20:07 other coding harnesses system prompts 20:10 that's in tokens. Yeah, 20:13 it turns out frontier models are heavily 20:15 RL trained to know what the coding agent 20:17 is. So why do you keep telling them that 20:20 they're a coding agent and how they 20:22 should do coding tasks, right? Um 20:27 YOLO by default. Um why is that? Um most 20:30 coding agent harnesses at the moment 20:32 have two modes. uh either agent can do 20:35 whatever it wants or agent um gets to 20:39 ask you do you really want to delete 20:41 this file do you really want to list the 20:43 files in this directory and so on and so 20:44 forth and there's different shades of 20:46 gray here but at the end of the day it 20:48 boils down to the user needs to approve 20:50 an action by the agent and then we are 20:52 safe and I think that's wrong because 20:54 that leads to fatigue and people will 20:56 either turn it off entirely yolo mode or 20:58 just sit there and type enter without 21:00 reading anything so I don't think that's 21:01 a solution container ization is also not 21:04 a solution if you're worried about 21:05 exfiltration of data and prompt 21:06 injections but I think that's the only 21:08 thing that you I think that's the best 21:11 basis uh compared to guard rates like 21:15 approval dialogues 21:17 it only has four tools read a file write 21:19 a file edit a file and bash bash is all 21:21 you need what's not in there no MCP no 21:23 sub aents no plan not no background bash 21:25 no built-in to-dos here's what you can 21:27 do instead for MCP use CLI tools plus 21:30 skills or build an extension which we 21:32 will see in a day. Uh, no sub agents. 21:35 Why? Because they're not observable. 21:37 Instead, use T-Max and spawn the agent. 21:40 Again, you have full control over the 21:42 agents outputs and inputs and can uh see 21:45 everything that's happening in the sub 21:47 agent. Interestingly enough, C code 21:49 spawn uh team mode now does exactly this 21:53 basically as well. No plan mode. Write a 21:56 plan MD file. you have a a persistent 21:58 artifact instead of some janky UI that 22:01 doesn't really fit into your terminal 22:03 viewport. Uh, and you can reuse it 22:05 across multiple sessions. Um, no 22:07 background bash, don't need it. We have 22:08 T-Max. It's the same thing. And no 22:11 built-in to-dos, write a to-do MD, same 22:14 thing. Or build all of this yourself the 22:17 way you like it. And this is what PI 22:18 allows you by being super extensible. So 22:21 you can extend tools, custom, you can 22:23 give the LLM tools that you define. I 22:26 think no other coding agent harness 22:28 currently offers that unless you fork 22:30 open code. You don't need to here. You 22:32 just write a simple TypeScript file and 22:34 it gets loaded uh automatically. 22:37 You can also write custom UI. Uh skills 22:40 are obviously in there. Prompt templates 22:42 uh themes uh and you can bundle all of 22:44 that up, put it on npm or git and 22:46 install it with a single command which 22:48 is very nice. And everything hot 22:50 reloads. So I develop my own extensions 22:53 that are project or task specific. um in 22:56 pi inside the project and uh as the uh 23:01 agent modifies the extension I just 23:04 reload and it immediately updates uh all 23:08 of the running code which is very nice 23:11 and in practice that means you can do 23:12 custom compaction I think that's one of 23:14 the things that people should experiment 23:16 more because all of compact of the 23:18 compaction implementations currently are 23:20 not good uh permission gates you can 23:22 easily implement them in 50 lines of 23:23 code and kind of cover what all the 23:25 other agent harnesses do if you want 23:27 that custom providers register proxies 23:30 or of self-hosted models don't care you 23:32 don't need me to do this for you can do 23:34 this and actually your clanker can do it 23:35 for you uh or overwrite any built-in 23:38 tool modify how readr edit and bash work 23:41 don't care I I have a version of readr 23:43 edit and bash that works through SSH on 23:45 a remote machine for me that took five 23:48 minutes to implement but it works uh and 23:51 you have full tui access so you can 23:53 actually write entirely custom UI in the 23:57 coding agent cloud code shipped slash by 24:00 the way it took 5 minutes for somebody 24:02 to replicate that in PI with more 24:03 features. Uh PI messenger I have no idea 24:06 what it's doing but apparently it's like 24:08 a chat room for multiple PI agents that 24:11 then communicate which then has custom 24:13 UI where you can look what they're doing 24:15 and yeah it just works. 24:18 Um, or Pi Mess. If you're bored, just 24:22 do play a game while the agent is 24:24 running, right? You can do that. Um, or 24:26 PI annotate. Um, open up the website 24:28 you're working on currently and annotate 24:30 stuff in the front end and give feedback 24:32 to the agent directly in line, feed it 24:36 back into the context, have it modify 24:38 the thing. Uh, or something I use is 24:41 file switch it. I don't want to switch 24:42 over to an ID or editor. I just want to 24:44 quickly look at a file that's been 24:46 modified. So all of this is extensions. 24:48 None of this is built in. And it takes 24:49 people usually a couple of minutes to an 24:52 afternoon to build all of this the way 24:54 they want it to. 24:56 Py web says also don't know what it's 24:58 doing. 24:59 Uh PI also comes with tree structure. 25:02 I'm not going to explain that. Just look 25:03 at pi.deaf. Um your session is a tree, 25:06 not a linear list of chats. So you can 25:08 basically do sub agents by read all the 25:10 files in the directory. Summarize this. 25:12 Go back to my my root of the 25:14 conversation. and take that summary with 25:15 me and do the actual work. Um, nothing 25:19 is injected behind your back. Agents, 25:22 skills, full cost tracking. A lot of 25:25 harnesses don't do this ear. Open code 25:27 does it not well. Uh, HTML export, JSON 25:31 format, headless JSON streaming, blah 25:32 blah blah. Does it actually work? Well, 25:34 terminal bench. Let me zoom in here. I 25:37 can't. This is amazing. Here's Pi 25:40 right behind Terminus 2 uh using Cloud 25:43 Opus 4.5. That was back in October where 25:46 Pi didn't even have compaction, right? 25:49 Uh demo time. Skipping that rate against 25:51 the clankers because they are breaking 25:52 open source 25:54 if you're associated with this guy's 25:56 project. Then you will have hundreds of 25:58 people coming from 26:02 OpenClaw to your repository and spam you 26:04 with clanker Philaw. Um, so I had to 26:07 invent a couple of measures. I invented 26:10 OSSification. So I just close issues and 26:13 PRs for a couple of weeks and work on 26:15 things on my own. Anything that's 26:17 important will be reported later on 26:19 anyways or in the discord. And then I 26:22 also uh implemented a custom access kind 26:25 of scheme where I have a markdown file 26:27 in the repository. If somebody opens an 26:29 PR without being in without their 26:32 account name being in that markdown 26:33 file, the PR gets autoc closed. I don't 26:35 care. First introduce yourself in a 26:37 human voice via an issue. Write an issue 26:40 that's not longer than a display long uh 26:42 because everything else is clankers lock 26:44 probably and once you did that I'm happy 26:46 to looks good to me you so you get into 26:48 that file and can now submit PRs to the 26:50 repository. All I'm asking is human 26:52 verification and Mitchell from Ghostly 26:54 then took this and took the uh uh built 26:57 a project called vouch which is more 26:59 easily applicable to your own open 27:00 source repositors and that is pi go 27:03 forth and try it