Vishal Banthia

@trq212 How caching would work here? Tools are defined before prompt. > But progressive disclosure is not just for skills, we also use it for tools. Some of our tools are ‘deferred loading,’ which means the agent must search for their full definitions using ToolSearch before using them.

We removed ~80% of the Claude Code system prompt for our newest models, this is what we've learned about writing system prompts, skills and Claude.MDs for them. x.com/i/article/2080…

@yacineMTB Curious to know how situation is different from Cloud Providers being the kill switch earlier?

We're quickly moving to a world where a small number of foreign corporations have a kill switch on all of your enterprises. Pretty terrifying

Use Fable to write Loops. Not to run Loop. #LoopEngineering

@deeeet Congratulations deeeet!

本日7月1日付でメルカリのCTO Japan Businessに就任しました． これまではSRE / Platform Engineeringとして技術基盤を，そして直近は越境事業において海外展開をエンジニアリングからリードしてきました．今後はCTOとしてJapan Business全体の技術戦略をリードしていきます．「組織と人」の面からAI-Native化を進めるCHRO/CAIOの @kimuras さ んと，「ものづくり」の面から連携しつつ，メルカリのAI-Native化をさらに進めていきます． AIを前提とした新しい開発体制へと移行すること，そしてAIが最大限に力を発揮できるように，これまで作ってきた技術基盤とアーキテクチャそのものを作り変えていくこと，この2つを軸にものづくりのAI-Native化を推進していきます．そして越境事業を軸としたグローバルへの展開とそのための基盤の整備も引き続きリードしていきます． よろしくお願いします!

started investing in a well structured wiki for better Agentic Search and its quite rewarding! And yeah, markdowns are totally fine.

My guess is this is more of acquihire for their expertise in understanding 3rd party API better. Anthropic need seamless integration with 3rd party app to make agents actually do useful things in enterprise market. MCP warmed up this but not a reliable and efficient solution. So acquisition is not for their SDK generator product.

Also someone please explain to me how these two narratives can coexist on here: SDKs are too hard, you need a flimsy startup to generate you SDKs from an OpenAI spec and it's the end of software engineering because clankers can do everything. x.com/mitsuhiko/stat…

For the last few weeks I've been down a rabbit hole trying to figure out what LLM inference actually costs. Given a model and some hardware, how much throughput can you squeeze, and at what price per token? To get my head around it I wrote a small simulator, then thought I'd add some visualizations on top. A few glimpses below. One caveat: the numbers come from a first-principles theoretical model, not measured benchmarks. It's a learning tool, not a production estimator. tokenomy.agenticloops.ai #LLM #Inference

@untraceable_the Write once and read many! If information is well structured, agent will have less round trips and also save thinking inference cost.

Genius play right here. Html costs more tokens They make more money smash like for more banger conspiracy theories

And it is not stopping. I had to disable the extra usage. @bcherny @trq212

Hey @claudeai I thought if I set my monthly limit then thats the max extra usage I will be charged. Here Claude Code decided to use 186% of the set limit. Who is going to pickup the tab? PS: I love Claude Code!

@gabriberton “Skilled hacker” was a rare skill. Now everyone has that skill. That is concerning

Super interesting take from one of the greatest hackers He says Mythos is not as good as they claim, because zero-day vulnerabilities are not that hard to find for skilled hackers I'm far from the hacking world but sounds reasonable Any thought?

If you can rip away the Bash tool from your agentic harness and can still make it useful then you will be at best contenders for enterprise adoption. Bash was a shortcut to show what an agentic harness can do but created too many security loopholes giving hard times for security team to review and approve.

False proxy. Same as number of commits or line added. Better metric would be combination of time saved and token usage efficiency. Time saved can be roughly evaluated using a user full session history and asking how much time saved to a model for the given task in the session.

Token usage is part of perf evaluations at Meta. This is just smart people (Meta only hires smart folks) hitting targets they assume leadership wants them to hit so they get that exceeds expectations (or above) rating + avoid below expectations (Perhaps this is Meta’s goal btw)

@theo Claude code is being designed for more general purpose harness. Also fuelling CoWork. Not only for coding. Probably thats why.

Can't stop thinking about how Claude Code is in LAST PLACE on TerminalBench for harnesses using Opus 4.6. There are TEN separate harnesses that use Opus better than Claude Code

@MythThrazz @ryxcommar @burkov You can run ccusage by npx ccusage@latest and see how much $ is saved because of caching.

@ryxcommar @burkov Still, the whole conversation is being sent over via the API, isn't it? And the KV cache (as each cache) is only good when/if it works.

Based on the leaked Claude Code source code, your CLAUDE.md file is re-injected on every single **turn** of the conversation.

𝗜𝘀 𝗖𝗹𝗮𝘂𝗱𝗲 𝗖𝗼𝗱𝗲 𝗮𝗻 𝗮𝗴𝗲𝗻𝘁 𝗼𝗿 𝗮 𝗵𝗮𝗿𝗻𝗲𝘀𝘀? Went down a rabbit hole trying to find the answer. Harness engineering is mentioned everywhere but the definition is not very clear, especially why we needed a new term in the first place. Then it clicked. It is about perspectives. 𝘾𝙡𝙖𝙪𝙙𝙚 𝘾𝙤𝙙𝙚 𝙞𝙨 𝙖𝙣 𝙖𝙜𝙚𝙣𝙩 𝙞𝙛 𝙮𝙤𝙪'𝙧𝙚 𝙩𝙝𝙚 𝙪𝙨𝙚𝙧. 𝘾𝙡𝙖𝙪𝙙𝙚 𝘾𝙤𝙙𝙚 𝙞𝙨 𝙖 𝙝𝙖𝙧𝙣𝙚𝙨𝙨 𝙞𝙛 𝙮𝙤𝙪'𝙧𝙚 𝙩𝙝𝙚 𝙗𝙪𝙞𝙡𝙙𝙚𝙧. An LLM is a wild horse. Raw power, unpredictable. A harness (bridle, reins, saddle) makes it controllable. Horse + harness = the agent. The product. The rider just rides. The harness engineer? They're not riding. They're building the system that makes riding possible. Harness layers stack. Claude Code provides the inner layer (agent loop, sandbox, tools, hooks, infrastructure). You build your layer on top (AGENTS.md, custom skills, verification gates). Someone else builds on top of yours. Same model. Different harness. Completely different experience. I wrote the full breakdown on AgenticLoops. First post in a series on harness engineering. Many more to come. Link in comments. #HarnessEngineering #AI #AgenticAI #PlatformEngineering Full Breakdown: agenticloopsai.substack.com/p/is-claude-co…

LLM only need two inputs. A prompt and list of tool. You understand that llm are trained on static data so it only knows that. Everything else that it needs to do meed to be added in the prompt (context engineering). Any action they need to take you provide that capability via tools. LLMs are smart and they will figure out a plan to finish the endgoal based on the available tools. You run a loop that basically keep calling LLM till the goal is met. In that loop it will use the available tools. This loop is an agent!

I think I finally understand what an agent is. It's a prompt (or several), skills, and tools. Did I get this right?

Your agents are only as reliable as your fundamentals. No shortcuts. We just added 8 new hands-on tutorials in advanced techniques module to our open-source project, each designed to strengthen one core skill you need to build agents that actually hold up. Hand crafted while keeping it simple with no framework bloating. Here's what problem each one solves: 𝟭. 𝗦𝘁𝗿𝘂𝗰𝘁𝘂𝗿𝗲𝗱 𝗢𝘂𝘁𝗽𝘂𝘁: Your agent returns messy, unparseable text and your downstream code breaks every other call 𝟮. 𝗦𝘁𝗿𝗲𝗮𝗺𝗶𝗻𝗴: Your users stare at a blank screen for 10 seconds waiting for a response and think the app is broken 𝟯. 𝗖𝗼𝗻𝘁𝗲𝘅𝘁 𝗘𝗻𝗴𝗶𝗻𝗲𝗲𝗿𝗶𝗻𝗴: Your agent forgets critical information mid-conversation because the context window silently overflowed 𝟰. 𝗖𝗼𝘀𝘁 𝗢𝗽𝘁𝗶𝗺𝗶𝘇𝗮𝘁𝗶𝗼𝗻: Your agent works great but your API bill is 10x what you budgeted and leadership is asking questions 𝟱. 𝗠𝗲𝗺𝗼𝗿𝘆: Your agent treats every conversation like a first date - no recall, no continuity, no learning 𝟲. 𝗥𝗔𝗚: Your agent confidently makes things up instead of using the data you already have 𝟳. 𝗠𝘂𝗹𝘁𝗶𝗺𝗼𝗱𝗮𝗹: Your agent can only read and write text while the real world runs on images, documents, and voice 𝟴. 𝗚𝘂𝗮𝗿𝗱𝗿𝗮𝗶𝗹𝘀: Your agent leaks PII, follows injected prompts, and hallucinates answers with zero safety nets github.com/agenticloops-a…

@jlongster Why agentic loop is running in a separate environment was something I was curious about. Thanks for clarifying that. Looking forward to see how session sync design pattern turns out.

OpenCode is about to get more powerful with remote sandboxes I showed a brief demo before, but here's a much more in-depth demo. it's not hard to add basic support for a remote env, but handling all the edge cases like when a remote env gets deleted is difficult. especially if care about good UX You never want to lose session data. so the choices are: run the session in your env, but run all tool calls remotely. that's too complex and painful. The other way is to just let the full session run remotely, but sync back all the session data in your env. We chose this path: we built a syncing system which logs all events in a way that we can always recreate your entire session. That means the remote env could get destroyed, but we can easily restore it. it also opens up other interesting ideas which we'll be exploring

@DanielLockyer Pattern still exists but the names are changed. It was a simple idea as most ideas are when you think retrospectively but revolutionary at its time. It created foundation for long running agents.

Funny how the "ralph loop" has completely disappeared from conversation