← Back to search

Episode 73: OpenClaw v2026.6.9, Hermes v2026.6.19, Claude Code 2.1.176 released; Poolside adds Laguna XS.2 and M.1

AgentStack Daily (हिंदी) · 2026-06-21 · 33 min
relevance 57 1251 words Episode page ↗ Audio ↗
Show full episode description
आज का AgentStack Daily: OpenClaw v2026.6.9, Hermes Agent v2026.6.19 और Claude Code CLI 2.1.176 सभी नए रिलीज़ के साथ आए। Poolside ने OpenRouter पर Laguna XS.2 और API के माध्यम से Laguna M.1 जारी किया। एंटरप्राइज़ टीमों को नए उपयोग विश्लेषण और अपडेटेड खर्च नियंत्रण मिले। Show notes: https://tobyonfitnesstech.com/hi/podcasts/episode-73/
✨ Episode Outline — click any point to jump to it in the episode
Problem solved
Builders need to track maturing agent stacks, new coding models, and production cost/security controls in one cycle.
Benefits
  • OpenClaw 6.9 richer Telegram delivery and auto plugin approvals
  • Hermes 6.19 adds background subagents and iMessage via Photon
  • Claude Code 2.1.176 lighter install, faster cold start
  • Laguna XS.2 pairs tools and reasoning in one endpoint
  • OpenAI enterprise spend controls and usage breakdowns
Use cases
  • Multi-file refactor agents using Laguna XS.2's cheaper per-token economics
  • Enterprise admins setting configurable spend ceilings on ChatGPT seats
  • Dataset Apps hosting custom HTML/JS dashboards beside agent data
  • Hermes dashboard full profile builder and revamped Skills Hub browser
KPIs / results
  • Laguna XS.2 context window 262,144 tokens
  • Laguna M.1 context window ~256,000 tokens
  • Basetten ~$1.5B round at $13B valuation
Tools / build
0:00 / 0:00
🌐 This transcript was automatically translated to English from the original.
I'm Nova, I'm Alay and this is AgentStack Daily. OpenClaw 6.9, Hermes Agent 6.19 and Terminal-Based AI Coding Agent Cloth Code 0.176 have all sent stable releases in this cycle. OpenClaw 6.9 added Richard, Telegram Delivery and improved KDEX Integration with Automatic Plugin Approvals, while Hermes 6.9 introduced background subs and extended Channel Support for iMessage via Photon. Today you'll hear about Unharness updates, the release of Poolside's Laguna Coding Models on Open Router and APIs, new Enterprise Spend Controls from OpenAI, and why Signal's President says your AI Chatbot is definitely not your friend. We'll also look at Bay Sittin's Billion Dollar Raise, a Vanity Search Engine that scores your presence within the model weights, and the historic failure of Export Controls through the lens of Anthropic's Mythos Model. The basic theme is to focus on production. How do we manage the cost of Agentic Loops? How to secure the model at their center? And how are the harnesses we use every day maturing to handle multi-step, multi-channel workloads? There were 3 stable releases in this cycle and these are shaping how Agentic Harnesses are being assembled right now. Open Cloche is DecimalNow, published June 21st Richard brings Telegram Delivery Channel Path now sends Rich HTML, Preserves Rich Markdown and Sticker Paths, and renders Progress Drafts and Command Output more faithfully It also Safely Normalizes HTML Tables and keeps Mentions and Spooled Handlers on the correct delivery path Agent Recovery is more dependable in this version, Retrize, Terminal Outcomes, and Session History Repair are now more Interrupted or Kodeks integration in OpenClaw has also been strengthened. It includes Automatic Plugin Approvals, GPT 5.3 Spark OAuth Routing and Remote Node Execution as a Dynamic Tool. These changes move Harness towards a more reliable App Server Teardown Process. Meanwhile, Hums Agent 6.19, which is called Maintenance Reach Release, extends Hums to new channels like iMessage and Raft Agent Network through Photon in Desktop App. Substantial new capabilities have also arrived. All Agents can now run in the background and the Image Generation Tool has learned to edit. Our Dashboard got a Full Profile Builder and the Skills Hub browser was completely overhauled. These updates provide a more robust backend for those using OpenAI Codex.141's Terminal Based Coding Agent. We also saw the Terminal Based AI Coding Agent Cloth Code.176 joining this trio of Stable Releases on the Stable Tag. This release has a lighter install footprint and Centered on a Faster Cold Start Path for Sheldriven Sessions These changes at the API and Runtime layer change what builders can configure and what they can relay on Defaults The practical implication is that Harness Plumbing becomes cleaner Open Cloth gives builders more Model Routes and Safer Auth while Hermes and Cloth Code refine the user and developer experience The part to keep an eye on is that these new Defaults under Real Workloads before flipping to production How to flow especially where Session History Poolside pushed the Laguna XS.2 on Open Router this week, which is the second generation model in their XS Size Class. Here's their Efficient Coding Agent Series. And here's Pitch Compactness Over Flagship Reasoning. It combines Tool Calling with Reasoning in a small footprint with a Context Window of 262,144 Tokens. This Context Size puts it right in the Long Context Tier for Agentic Coding Workloads where Large Code Bases or Tool Traces are nearby. The most relevant details for builders are that Tool and Reasoning are a single integrated pair A single Endpoint exposes both Tool Invocation and Reasoning which simplifies Agent Loop Orchestration You don't have to route requests to two different models to handle Planning and Execution Access Size Class Completeness Indicates Target Audience Focusing on cost-sensitive runs If you're running a Multi-File Refactor Agent The per-token economics here may be more attractive than the Flagship Class Reasoning models For Agent Builders This release widens the Compact Tier on Routers Sits here alongside other Compact options that are used for cheaper Classification or Initial Planning Passes Next up we need to see actual Latency Benchmarks on real Coding Traces We also want to see if Poolside eventually releases Tool Calling specific Rate Limits or Pricing Tiers to further differentiate the XS Line from its larger Compact models This refresh of the tier is important because it pushes not only the limits of raw quality but also the limits of cost. If your Agent Stack relies on a lot of small Iterative Calls then the Second Gen Compact Model with a huge Context Window is an important integration target. This allows for more complex multistep workflows without the latency hit of the larger Flagship Model. Moving to the other end of the spectrum, Poolside has also released Laguna M.1 via API. Here is their Flagship Coding Agent Model which provides complex Soft Optimized for ware engineering practices Like its younger brother, it supports Tool Calling and Reasoning with a 296 token Context Window. At the system level, this change is reflected in the API surface and runtime behavior against which Agent Builders integrate. Laguna M.1 is a release targeted at the heavy lifting of agentic coding. Think architectural changes, complex bug fixes, and large-scale refactors that require a deep understanding of the entire stack. Prathmix Rote of this Model Documentation includes specific Deployment Notes and Change Log references that builders should review. It's worth tracking how MA behaves under heavy production loads because flagship coding models often face unique challenges with consistency over long traces. This matters now because of how fast the Agent Stack is growing. Changes in this layer determine which workflows are reliable and which still remain fragile. The practical question for builders is whether Laguna M.A can replace their current flagship default coding tools. Early evidence suggests it is a strong contender for those building autonomous or semiautonomous engineering agents. We should keep an eye on follow-up releases and independent benchmark results. As surrounding tooling like SDKs and security review frameworks start to gain support for M.1, the barrier to switching will go down. For now, focus on what to expect from your agents. OpenAI Chat is validating its performance against specific coding patterns OpenAI Chat GPT is introducing new spend controls and usage breakdowns for Enterprise This is a direct response to the needs of organizations that want to manage costs and scale their AI deployments with more confidence These controls are applied at the management and API level, allowing admins to customize how to configure and deploy seats across a large company It includes more detailed cost breakdowns and configurable ceilings on spend This tool for builders and platform engineers includes the ability to set It's about visibility as agents spread across the company It's important to know which departments or projects are driving the usage The updated visualization here gives a clearer view of how tokens are being spent which helps predict future budget projections It also helps identify inefficient agent loops that can burn through the budget without delivering equivalent value These spend controls change what the enterprise stack defaults to. Admins can now set limits in advance, rather than having to respond to high bills at the end of the month It's worth tracking how these controls impact agent performance, especially if the ceiling is reached mid-season Builders need to ensure their agents can handle graceful declines or clear notifications when budget limits are reached This move to Open AI reflects the maturing of the enterprise AI market Cost management is no longer a consideration here Production We expect other major providers to follow suit and come up with similar enterprise-grade visibility and control features as the focus shifts from ease of use to operational efficiency. A recent study argues that 30 years of US export controls on encryption and cybersecurity software have largely failed to slow their spread. This historical context is now entropic. The main problem with which Mythas is being applied to the cyber security model is that due to the fact that software has always been leaked, forked and reimplemented regardless of ownership, the historical mechanism that failed was to treat source code or compiled binaries as controlled artifacts. The challenge for Mythas is even greater. The question is whether the model is weighted, trained, computed or hosted. Inference API can be effectively controlled. Unlike compiled binary, it is difficult to limit the capabilities of a model to a specific artifact. This makes the manipulation quite complex for regulation. The visualization shows that the propagation mechanism for AI could potentially be a cryptographic tool. Code folk and overseas re-implementation will override any regulatory framework The key takeaway for builders is that frontier AI cybersecurity capabilities will likely be globally accessible You should assume that both defensive and offensive AI security tools will be available for a variety of tasks This makes export classification a moving target for compliance teams If you're building with or against Mythos, the model The classification of weights and inference endpoints remains a point of ambiguity. We should keep an eye on any new rulemaking from the Commerce Department regarding frontier model weights, as well as whether Anthropic publishes a usage policy that attempts to address these regulatory questions in advance. This fight over Mythas will likely set the policy framework for the entire AI security sector. AI Inference Startup Basiton's $1B Funding The round is reported to be finalized at a valuation of $13 billion. This funding comes just a few months after their previous mega round and marks a major shift in the market. Inference is no longer seen as a feature of the Model Training Lab but as a separate infrastructure category in its own right. Dedicated funding is going into companies that focus specifically on the serving side. The technology here is the Inference specific serving stack. Basiton's differentiator is to expose these NOBS as a managed service for engineering teams. Builders get more control over request routing and optimization for bursty or long context traffic, rather than relying on an opaque API endpoint from Model Lab. This huge funding signals that the inference layer is becoming a buyer's market. There is now real competition between specialist providers like Basiton, Fireworks and Together. For builders this means that the default choice of inference provider is now clear. You must clearly evaluate the tradeoffs between cost per token, latency, and the level of customization available. We'll be watching here to see how the Basiton Hyperscalar Inference API positions itself against the likes of Bedrock or Azure AI. The pace of innovation at the Serving Layer will likely accelerate as more capital enters the space. There's good news here for Agent Builders looking for reliable high-performance infrastructure to power their Production Loop. Dataset has launched a new plugin called Dataset Apps that allows users to host custom HTML and JavaScript applications directly inside Dataset. These are self-contained applications that can take advantage of data stored in a Dataset instance. The launch announcement outlines the why behind the move. It focuses on the need for more flexible ways to visualize and interact with data without the need for a separate hosting stack. This UI for Agent stack builders An interesting development in the layer is that by hosting custom HTML within the data tool itself, you can create a tighter feedback loop for agents that are interacting with the database. This makes it easier to deploy internal tools and dashboards that need a specific front end but need to be closely coupled with the data source. The change here impacts the API and runtime level, which impacts the way you configure plugins and deploy your custom apps. This change means that builders can access the dataset for data driven agents. As a more pooled platform you can rely on, it's important to track how the plugin handles different workloads and how it scales with complex JavaScript applications. If you're building tools for data exploration or agent monitoring, this can significantly reduce your infrastructure overhead. Here we'll look at how the community adopts data set apps and what kind of custom interfaces emerge as more builders experiment with it. We'll see the agent human collaboration surface. e will see new patterns built directly on top of the data they are processing A relatively slow period for AI news has been noted in recent newsletters No big model drops or agent framework releases dominating the cycle This breathing room is often a sign of calendar-driven release coordination AI engineers or A conference is coming up Many of the big lass and maintainers are submitting their reveals for that event This drop before the conference is a standard pattern in the ecosystem This calm for builders window is actually quite useful. It gives you an opportunity to consolidate notes on current agent frameworks and stable API versions. You can commit to your current stack for a few days without the high risk that a breaking change will disturb your local environment. This is a good time to focus on perfecting existing workflows and solidifying your session handling and tool integrations. The thing to look for here is the opening keynote. Historically, model providers and agent runtime maintainers used these stages to ship reference implementations and pin version releases. These announcements are usually spread across documentation and repositories within a few hours of the talk. The quieter the runup, the more impactful keynote reveals are. Expect a whirlwind of activity as soon as the conference begins. Enjoy the stability for now and prepare your stack for the next wave of updates. Changes in the post-conference shipping cycle are usually very rapid this fall. Meredith Whitaker, President of Signal, recently discussed the trend of positioning AI chatbots as companions in an interview. Their message was clear: These are not your friends. No, they are conscious beings and not sentient interlocutors. This critique directly targets vendors who use relational language in their onboarding flows, persona system prompts, and conversation design. This intervention is especially relevant when using agentic coding tools. AI and support bots are becoming embedded in our daily workflows The line between tool and teammate is being blurred by anthropomorphic design choices Witteka argues that framing models as peers creates false expectations about their memory and intent For builders, this means your agent's product copy and system prompts are now part of the trust surface Anthropomorphic framing in Persona prompts or UI greetings may be a privacy red flag for some users Signal The focus on privacy gives this critic a lot of weightage in developer circles. As a builder, you'll want to look at whether your agent persona is helping or hurting user trust. Relational copy may ultimately become a deal breaker for privacy focused enterprise buyers. I'll see if the major model providers change their mark philosophy about relational framing in their developer documentation. If enterprise buyers start flagging anthropomorphic copy, we may see a further return of the utilitarian and transparent agent persona. How we frame agent-human interactions is becoming a key design decision. A new service called in the weights has launched that is pitching itself as a vanity search engine for the AI age. Instead of indexing web pages, it assigns users a score based on how prominently their identities appear within the parameters and training data of frontier AI models. It uses the models as a search index that provides a measure of how well the AI is performing. The basic mechanism behind this involves repeatedly running inference probes against hosted LLMs. The architecture model uses an evaluation harness to query identity mentions at checkpoints. It then aggregates hit rates and frequency counts into a single composite score. The model reframes visibility as a measurable metric that first actually measures product-based survivability. Did not exist in the form of face AI This could become a new category of user facing metrics for product shippers It surfs a different kind of signal than traditional search ranking The question for builders is whether the scoring methodology will be standardized or will remain opaque Reproducibility The question is whether this score will become a real signal for marketing or will be just another vanity number I will see if a probing protocol is opened for this service If so, then we will make different model versions You can see new tooling being created around tuning your presence in the Internet This is an interesting look at how we can measure the influence of individuals and brands in an AI-filled information environment The first one on the radar is fast NCP from the prefect team This is a pythonic framework designed to create servers and clients using the model context protocol It allows developers to expose tools and resources to LLMs with minimal boiler plate If you have internal python functions that you want to use in the cloud fast NCP provides a standard way to wrap them as typed tool calls. Here's a great replacement for ad hoc CLI wrappers. Next is the model context protocol for beginners repo from Microsoft. It's a comprehensive curriculum walking developers through the fundamentals of the protocol. What makes it useful is the cross language examples across .NET, Java, Rust, and TypeScript. It provides an obvious sin for wiring agents in non-python services without having to rewrite everything. You can choose a lab in your preferred language and start porting your existing internal tools to the MCP server. Finally we have Unity MCP. This project builds a bridge between AI assistance and the Unity Editor. It exposes tools for asset management, scene control and editor automation. It makes the Unity Editor API reachable through standard NCP tool calls. It makes the Unity Editor API reachable through standard NCP tool calls. For builders on space, it turns asset pipelines and scene edits into agent driven steps. You can orchestrate the entire build and test loop through your agent stack. There are two important entries from the poolside in this cycle's model check. The first is Laguna Axis.2. Now available on Open Router. As we discussed earlier, this is the second generation compact model in their axis class. It has a context window of 262,144 Tokens and it combines tool calling with reasoning if you If you are looking to optimize cost and latency in your agent loops, this is a prime candidate to evaluate against your current defaults. The second selected model is laguna m.1. This is the flagship entry from Poolside. It is also available via API. It is optimized for complex software engineering tasks and supports similar massive context windows. Its capabilities are aimed for higher tier agentic coding workflows. We recommend routing some coding agent sessions through both m.1 and xs.2. To see how they handle your specific code base compared to your current primary models. Other models were reviewed in this cycle but were not selected for full coverage because they did not represent a new major provider release. We continue to track the long context and tool calling performance of all the major models on the market to ensure that your agent stack has the best possible foundation. For our local spotlight we're looking at olama version 0.30.10. The update adds accelerated support for Coher's command A and North Family of Models on Apple Silicon. It also upgrades the embedded lama engine to build 9672. For builders who want to keep their workloads on one device, this is an important update because it makes local inference of model families on a series Mac possible without the need for CUDA. If you're working on a series Mac, you can pull the command A or North model and pull the token per second back to your previous level. ulama bui Can benchmark against ld This release tightens the loop for agents who need to work in high privacy or offline environments The performance gains from MLX acceleration are noticeable and help make local agents much more responsive In our first extra, we look at the ambitious ambitions of billionaire Mukesh Ambani, who wants to weave AI into every call, app and home in India The technology cone here on a device and age inference in the telecom carrier stack embedding which has over 500 million subscribers This represents a huge scale for the network resident AI surface Next up is the story of the Allbirds CEO who launched a new AI business with a plan but no employees This raises the question of how a single founder company can operationalize model selection and shipping by relying entirely on an outsourced or vendor provided inference stack Here's an interesting look at the company of one model run by high tier agents Finally a controversy Whether ASM LK top chipmaking tools have found their way into China The story hinges on how lithography export controls verification and the physical location of such restricted tools are tracked highlights the friction between commercial logic and national security regulations in the high-end hardware space that powers AI From today's stories the stable releases of opencloth 6.9, harms 6.19, and cloth code 1.176 make your agent stack the default tool you can rely on. change that Poolside's laguna axis.2 and m.1 give you new options for coding agent loops with large context requirements openAI's new enterprise controls provide the visibility needed to manage agent spend at scale The historic failure of export controls shows that defensive and offensive AI capabilities will be globally accessible, making security a baseline assumption The massive race for base setting shows that the inference layer is now a separate This is an unending and highly competitive market. Finally, remember that the choice of your agent persona and copy can impact user trust, especially given the privacy concerns raised by Signal. You can find all the details and all the links in the show notes toby on fitnesstech.com. Agent Stack Daily Thanks for listening we'll be back soon.