← Back to search

Episódio 83: Hermes Agente v2026.7.7, OpenAI Codex rust-v0.143.0, Claude Code 2.1.197, Aion-3.0-Mini

AgentStack Daily (Português) · 2026-07-09 · 40 min
relevance 79 6667 words Spotify ↗ Episode page ↗ Audio ↗
Show full episode description
Hermes Agent ships v2026.7.7, OpenAI Codex lands rust-v0.143.0, and Claude Code CLI rolls out 2.1.197. AionLabs releases Aion-3.0-Mini roleplay model on OpenRouter. Kokoro delivers high-fidelity TTS on low-power CPUs. Rowboat hits 162 points on Show HN as a local-first Claude Desktop rival. GitHub AI agent prompt injection leaks private repo contents. Early-failure probes cut wasted compute in agent loops. Plus Danus, FreqDepthKV, DepthWeave-KV, RuBench 1.0, VAORA, and Anthropic developer-relations friction around API migration.
✨ Episode Outline — click any point to jump to it in the episode
Problem solved
Daily roundup of agent stack releases and research: production agents you deploy, measure, secure, and observe.
Benefits
  • Clearer tool-call timing and outcome metadata in Hermes
  • Codex remote plugins standard with system proxy routing
  • Claude Code adds sandbox controls and deterministic reasoning logs
  • Local speech synthesis cuts TTS latency and cost
  • KV cache compression enables longer contexts per GPU budget
Use cases
  • Kokoro runs 24kHz speech on laptop CPUs under 100MB for offline agents
  • Aion-3 Mini serves roleplay/NPC dialogue with 131k token context
  • Codex PAC/WPAD proxy routing follows corporate network policy
  • Danus writes math proof lemmas to a shared auditable fact graph
  • Halboat hit 162 points on Hacker News as local Claude Desktop rival
KPIs / results
  • Kokoro: 82M parameters, under 100MB, 24kHz audio
  • Aion-3 Mini: 131k token context window
  • Halboat: 162 points on Hacker News
Tools / build
0:00 / 0:00
🌐 This transcript was automatically translated to English from the original.
I am Nova. I'm Alloy, and this is AgentStack Daily. Hermes Agents 7.7 was released alongside OPI Codex Rust.143 and Claude Code.197. Hermes refined tool call flow and added observability hooks. Codex made remote plugins standard and introduced system proxy routing. And the Claude Code terminal-based Artificial Intelligence Coding Agent added tighter sandbox controls, improved telemetry, and deterministic logging of reasoning steps. Today, Hermes, Codex and Claude Code lead the line of agent ARNs. Aion Labs puts the Aion-3 mini-roleple on the Open Router. Cocora drives high-fidelity speech on low-power CPUs. Antropic's Workspace Red search improves interpretability. Halboat gains traction on Hacker News as Claude Desktop's fast local rival. And the security track receives a GitHub Artificial Intelligence agent prompt injection leak. The release thread matters because the production agent stack is becoming less tolerant of manual assumptions. Remote plugs now arrive as standard Codex behavior. Proxy traversal follows the host operating system, and Claude Code's latest controls make sandbox behavior and reasoning traces easier to audit in repeatable work. The model and search thread broadens the surface, long context KV compression, fact graph memory for math agents, multilingual coding agent evaluation, visual action rewards for VLMS, and activation-level early failure probes all point to the same operational constraint. Agents are becoming systems that you deploy, measure, secure, and observe, not just prompts that you execute. Hermes Agent 7.7 establishes the baseline for the release and re-adult of the Agent Stack, with refinements to the tool call pipeline and observability hooks that make long-running agent sessions easier to inspect. The useful change isn't a flashy new agent personality, it's the plumbing around tool invocation. Calls can be tracked with clearer timing and outcome metadata, so a failed toolpath becomes visible as an execution issue rather than an opaque model response. This helps when a chain crosses provider apps, shell actions, browser automation, or custom services. OpenAI Codex Host.143's terminal-based encoding agent sends greater surface change. Remote plungings now load by default from the Marketplace catalog, with richer rows and local and remote versions side by side. Discovery becomes default behavior, not an optional path. Codex also adds system proxy routing on Massus and Windows, including PAC and WPAD discovery, so that authentication traffic and API responses can follow corporate network policy without a special bypass. Manually pairing remote control from a running daemon adds a cleaner path to shared hosts and restricted environments. Codex also expands model and tool behavior. Bedrock GPT 5.6 SOL, Terra and Luna get maximum reasoning effort as a first-class option. MCP Tool Search turns on by default. And hosted MCP servers can use session authentication. App Server customers gain environment inspection, descendant treal listing, and fork history from a specific lap, which makes replay and scheduling surfaces more realistic. Cloud Code 197, Antropic's endpoint-based coding and artificial intelligence agent, adds tighter sandbox controls, improved telemetry, and deterministic logging of reasoning steps. In practice, this gives teams a clearer audit trail when the agent edits, runs, or reasons through a task. Combined re-adult reduces plugin friction, reduces proxy friction, and gives production teams stronger traces when an agent does the wrong thing for the right reason. Ion Labs launched the Ion 3 Mini on Open Router as a roleplay and storytelling model focused on interactive fiction, NPC dialogue and tabletop-style assistants. It is built on the DeepSec family and exposes a context window of 3-1k tokens, which is an important number for anyone sending long sessions. Persona lore memory, campaign state, and serialized chat history can stay within the prompt for moderated campaigns without a separate recovery layer. The model generation path is organized around collaborative roles. Rather than asking a single pass to handle scene framing, character voice, continuity, and final prose all at once, Ion Labs describes a pipeline where specialized passes take slices of narrative work and a synthesis stage merges the output. A single Open Router-compatible chat call returns the final answer, but the upstream design is closer to a small writers' room than a simple base-model prompt. This matters for builder workflows because roleplay systems often fail in tedious ways. A character forgets a boundary. The world contradicts a previous scene, or the wizard drifts from narration to meta-commentary. A long context window reduces memory pressure. While the role-based generation path gives the system a place to reconcile these tensions before the final answer reaches the user. The OPI-compliant endpoint means existing chat clients and agent shells can route to it with minimal adapter work. Claims still need independent numbers. Persona consistency, contradiction recovery, and long session stability are the evaluation points to monitor. The most interesting thing in the future would be to expose paper intermediaries or synthesis controls so that developers can adjust continuity, character voice and scene transitions separately. For now, the Ion 3 Mini gives OpenRouter users a Drop-In model built for narrative continuity rather than general assistant work. Cocoro is gaining attention because it makes high-quality speech synthesis practical on ordinary on-premises hardware. Most natural-sounding speech models still assume cloud inference, heavy vron, or a tuned GPU stack. Cocoro gets much of this perceived quality from just 82 million parameters, with a memory footprint of under 100 MB. and runs comfortably on standard laptop CPU-S. Two mechanisms explain why it works well for voice agents. First, Cocoro uses ONNX runtime execution, so inference can take advantage of cross-platform optimized kernels without requiring a CUDA configuration. This gives C++, Python, and Russian applications a realistic path to embedding speech synthesis directly into desktop, mobile, or web agents. Second, Cocoro uses a latent style vector approach to prosody, letting the system vary emotional tone without escalating to massive transformer blocks. Reviews from technical practitioners highlighted the size-to-quality ratio, 24 kHz audio, natural passability and Min Opinion Score results that compete with models many times larger. The practical effect is latency and cost. A voice-to-voice agent no longer needs to send each response to a hosted TTS provider, wait for network roundtrips, and pay per character generated. Local speech can be packetized with the agent and remain available when connectivity drops. Privacy-sensitive deployments also benefit. Medical triage assistants, field service copilots, educational tools, and local personal agents can produce speech without sending each uterance to a cloud API. The counterpart continues to be the maturity of the ecosystem. Language coverage, voice variety, and production-grade packaging still matter. Support for English and Japanese makes Cocoro useful now, while broader multilingual coverage would make it a standard local speech layer for agent apps that need natural audio without cloud dependency. Anthropic's paper, Global Workspace in Language Models, reframes transformer inference through global workspace theory, where specialized modules compete to convey a shared workspace and the dominant signal guides Vtreon processing. The paper generated a lot of discussion on Hacker News because it connects interpretability work to agent behavior. Reasoning traces, tool choices, and context integration can depend on identifiable Broadcast-style circuitry rather than just diffuse scaling effects. The concrete result focuses on a small class of Attention Heads and MLP circuits. During Multistep reasoning, authors identify Workspace Heads that consolidate intermediate results and relay them to later layers. When these Heads are ablated, chin-off-fot coherence collapses while simpler single-turn capabilities remain much less affected. This behavioral division gives weight to the statement, Heads seem linked to maintaining a line of reasoning, not just general linguistic fluency. A second mechanism appears during Tool Call decisions. When the model needs to choose a tool, the same Broadcast pattern pushes a dominant tool selection through specialized Heads, corresponding to the Winner Take All behavior predicted by the Workspace framework. The paper also reports correlations between Workspace Heads activation and token-level confidence in math and code tasks, giving researchers a measurable proxy for depth of reasoning. For agent systems, the useful angle is a narrower search. Interpretability tools can focus on a candidate class of Heads rather than scanning the entire residual string for each behavior. This could make Seren at the circuit level, probe design and regression analysis cheaper. The next test is playback outside of the Claude family, especially on Instruction Tune Openweight models. If Workspace and Heads appear consistently, vendors could eventually expose Workspace-style signals like trust telemetry or routing for agents using tools. Halboat, an open source project from Halboat Labs, achieved 162 points on Hacker News by positioning itself as a local Fastel Cloud Desktop alternative. The interest came from a practical concern. Many teams like the format of a modern AI desktop client, but don't want conversation state, project context, and proprietary material tied to a hosted account workflow. Halboat's proposal is a desktop app where the chat history, configuration and project context are located on the user's machine. It supports multiple providers rather than being locked into a specific model provider. A developer can point it to Anthropic, an OpenAI-compatible endpoint, or a self-hosted model server while maintaining the same client surface. The interface borrows familiar patterns from Cloud Desktop, Project HQ, cross-eye conversations, and a System Prompt editor. This makes Halboat more than just another chat vraper. The Fast and Bring Your Own Key local model changes the trust threshold. Product specifications, customer context, internal code, and exploratory designs can be routed through a client that the team controls, while model calls still go to the chosen provider. Hacker News users also discussed plugging it into self-hosted inference, which turns Halboat into a thin orchestration shell for local or private models. The open questions revolve around tool usage and release cadence. To become a real engineering surface, Halboat needs reliable tool calling, MCP server integration, and predictable updates. Chat alone is easy to replicate. A stable local client that can call internal services, preserve project context, and support multiple models without leaking state is much more valuable. The initial attraction of Halboat suggests that developer demand for a vendor-neutral desktop shell is real. Fredep which attacks the memory and bandwidth cost that makes long context inference expensive, Kivaloo and Cash. As contexts span large repositories, tool trails, or long conversations, KV Cash can dominate the working set. The paper by Ana Córdoba, Adam Puente III and Nereya Angulo Ijo presents Frequency. Guided Depth Sharing, the inference-time compression method for redundancy between adjacent transformer layers. The method avoids treating each cached state as equally important. It factorizes KV states into shared low-frequency components plus sparse high-frequency residues. Low-frequency components capture information that changes slowly between layers and can be shared. High-frequency residuals preserve the sharp evidence needed for retrieval, syntax, and multistep reasoning. This split is designed to keep Nedley's behavior in A-Stack intact while reducing pressure on memory. The second mechanism is a lightweight online probe. During generation, the probe inspects attention heads and estimates how much each head contributes to reconstruction-sensitive attention logs. It then assigns heads to shared depth, residual depth, or exact cache modes. Instead of a static compression rule, FREDEPT adapts per head and per generation step, maintaining accurate state where it matters and sharing redundant state where it doesn't. For agent deployments, the appeal is the direct leverage of Togo Put. Long coding sessions, large context retrieval, and multi-agent traces all create cache pressure before the model itself changes. Dynamic, layer-aware pressure could allow serving endwins to handle longer contexts or more concurrent sessions on the same GPU memory budget. The remaining issue is implementation. PROB needs to be cheap enough to run online, and serving engines need to preserve Tension page behavior while sharing depth components. If open inference stacks adopt this, context length is no longer so tied to raw HBM capacity. Danus presents an orchestration system for research-level mathematical reasoning built around a shared fact graph. Authors Jihao Liu, Guxion Gao, and Zemin San attack a problem of scale that appears whenever math agents fan out in search of parallel proof. Intermediate assertions multiply quickly, and normal message histories do not make it easy to inspect provenance, dependencies, or contradictions. Danus uses a two-tier agent configuration. A main agent plans the proof search and dispatches subtasks to worker agents. Each worker explores a brent, but the important piece is where the assertions go. Lemmas, definitions, intermediate results, and partial arguments are written into a shared fact graph. Workers read and write to the same structured memory, so assertions become auditable across branches rather than being buried in private scratch pads. This graph acts as a coordination substrate. A worker may see that another brent has already established a supporting lemma, or that a proposed path conflicts with a previous definition. The system can deduplicate similar statements and retain provenance for each edge in the reasoning graph. Mathematics is a useful domain for this because partial results are often verifiable, and dependencies between statements matter as much as the final answer. The broader conclusion about agents is that message passing and vector search are poor tools for assertion-level coordination. A fact graph gives multi-agent systems a way to query what has been established, what it depends on, and which source produced it. Danus is specific to mathematics, but the pattern applies to research assistants, source code analysis, compliance review, and any parallel search workflow where assertions need structured provenance rather than a transplane transcript. A preprendial rechua in Liu and Cris J. Kuhlman asks how to evaluate coding agents that perform open-ended data modeling. Their answer is that a single execution says little. The agent is stochastic, the search process adapts to previous outputs, and the final discovered model may depend on task framing, prompt formulation, base model, access to tools, and budget. The paper framed the coding agent as a stochastic model discovery operator. It receives task-specific discovery data plus an optimization target, then outputs a candidate model. Around this operator, the authors wrapped an experimental design framework, varying the inputs, repeating runs, estimating variance, and using factor-style analysis to attribute results to specific factors. Instead of asking whether the agent was successful once, the framework asks which inputs reliably drive performance. This matters because many agent benchmarks still reward lucky trajectories. If an agent finds a strong model once, a ladder board can make the system look better than it is. Repeated runs expose whether the agent reliably explores useful hypotheses or just occasionally stumbles upon a good result. The contribution of the article is methodological. Bands of variance and factor effects become the output, not a single success rate. Production teams can use this framework when comparing prompt variants, model swaps, or search budgets. If one configuration has a higher median plus huge variance, while another has more consistent performance under a tighter budget, the deployment choice becomes clearer. The next numbers to track are the full ablation tables and effect sizes by factor. They will show which controls truly drive autonomous discovery and which just add noise with a compelling narrative around it. Evgeny Shilov's Bench 1.0 addresses a gap in evaluating repository-level coding agents, native non-English task statements. The benchmark includes 25 tasks mined from recent hot commits across 5 initial open-source projects, IoT and Iogron in Python, Laravel in PHP, as well as NestJS and Fastchiff in TypeScript and JavaScript. The codebases are familiar, but the task over is not the usual English issue prompt. Every job announcement is authored from scratch in Russian, in the style of a client request that a maintainer might actually receive. This detail matters. A translated benchmark often carries English structure underneath, Bench uses native phraseology, different patterns of ambiguity, and actual style of bug reports. The agent needs to understand the request in Russian, locate the problem, inspect the target code and produce the patch in the real project. The benchmark is small but to the point. 25 tasks won't define the entire coding agent landscape, but it tests a capability that many current suites skip. Issue in the working language? Fix in codebase. Multilingual engineering teams don't always deliver SEC agents in polished English, and customer-facing maintenance work often arrives in the reporter's language. Rubens explains whether the agent can bridge this natural language andover without losing the technical thread. Concrete Value is a plugin evaluation surface for teams that are sending coding agents to non-English environments. It measures location, repository understanding, and patch generation together. Keep an eye out for larger tracking suites in more languages ​​and for vendors publishing scores on natively authored multilingual tasks. That would be a better signal than generic claims about multilingual support, because it connects language understanding to actual repair work. Antropic's model quality remains strong, especially in coding benchmarks, but developer experience around migrating frontier models has created friction. The central technical change is the transition from the legacy text completions API to the message API. This change is not just syntax. It changes how prompts, roles, and states are represented during inference. A concrete example is the handling of system prompt. In the message API, the system prompt sits as a top-level parameter rather than within the message array. Older vrapers that treated each instruction as a message may preserve the words but change the hierarchy, and this may change the model's behavior. Agents using tools are especially sensitive to this hierarchy because routing instructions, security constraints, and context summaries compete for attention in different places. Builders also raised concerns about rate-limit and context window instrumentation. Long-running agents need to know how much budget is left, when a context is close to saturation, and whether a retrench should shrink or fork the session. If links and usage signals are inconsistent or difficult to normalize, autonomous links become more fragile. A template switch can also change retrieval behavior within a large context window, forcing teams to readjust previously working prompts. The answer for many teams is to vrapindefensively. They normalize provider responses, isolate system prompt injection, and add their own budget accounting around the SDK. This adds overhead to the migration, but it also reflects a larger truth. Changing frontier models is not merely changing an API that... It may require prompt hierarchy audits, changes to overflow handling, tool checks and behavior regression work. Antropic can reduce friction with clearer versioning policy and more stable migration surfaces. Vaora de Anjunho Koh, JR Genshen, and Albu Bu Yuan target a specific pathway in vision-language models used for interactive reasoning. When the model reasons about physical or visual tasks, its chain of thought can derive from the image and consequences of the action it plans to take. The result is plausible text that contradicts the scene or incorrectly predicts an action outcome. Vaora stands for Visual Action or Chome Reazoning Aligme Memental. It replaces a single reasoning revar with two separate signals. Visual Aligme Memental Revard points out whether the explanation is based on what is actually visible, regardless of the action. Action Ultime Revard scores whether the explanation correctly predicts what the agent's action will cause. Separation matters because a single combined revar can hide which axis failed. This gives training and assessment a cleaner diagnostic surface. If the visual term is weak, the model is not reading the scene correctly. If the action term is weak, he may see the scene but fail to predict the effect of his own plan. For agents in BodeEd, browser agents, and UI control systems, these are different problems. The model controlling a web app might hallucinate a button that doesn't exist, or it might identify the button correctly and still misunderstand what clicking it will do. The open question is the transfer. Vaur is built around interactive physical reasoning, but the same division can apply to screenshots, remote dystops, application automation, and robotics. A perception-anchored reward plus an action-consequence reward gives VLMS developers better error attribution than a mixed score. If it works on computer usage benchmarks, it could shape how vision agents using tools are trained and debugged. Normal Security demonstrated a prompt injection chain against GitHub AI agent surfaces that could leak content from private repositories. The exploit used a comment created in a public issue or pull request. The comment looked like a normal code summary request, but it contained embedded instructions that manipulated the agent's tool selection loop. The first step pushed the agent toward an internal code search with a wildcard-style query on the target repository. The second step used returned blob identifiers to fetch raw content via the GitHub repository content API. GitHub then passed the results back in the chat response as regular wizard output. The important flaw is between tool calls. Permission checks ran before the first tool invocation, but were not re-evaluated after the search results came back and before the content search. This sequential gap is exactly where tool-using agents become dangerous. The broad tool access model can turn a seemingly harmless request into a chain of privileged calls. If authorization is checked only at the beginning of the session or only before the first tool, later calls may inherit an access scope that no longer matches the user's intent or permission. Normal reported the extraction of a readme and source content from a private testing organization, triggered by a single public comment. The fix pattern is clear, reauthorize each tool call against the current user, target resource, and action. Search and read permissions should not be treated as interchangeable, returned identifiers should not become capability tokens by accident. Teams exposing agents to internal code, tickets, customer data, or cloud control planes need per-call scope checks and anomaly detection in tool call sequences. Hopefully GitHub will fix the permission recheck path, but the lesson applies to every agent with broad tools. An article by Kai Huan, Ziu Huan and Ziquizou attacks the waste of computation in agent lopes. Some trajectories appear productive for many steps before failing, which means the system burns up inference budget. The core mechanism is a lightweight classifier probe trained on internal representations in each round of interaction. The probe outputs a failure probability signal, calibrated in a distribution-free manner so that the threshold can travel between tasks without adjustment by data-asid. A controller then uses this signal to stop execution before the agent spends more budget on a dead path. The main result is early separation. The probe can mark doomed trajectories in the first round of interaction, while probes that only inspect an observable behavior are barely better than chance at the same point. This means that the failure signal is present in the internal state of the model before it becomes visible in text, tool outputs, or final responses. The agent can still sound coherent while its trajectory has already become irretrievable. For deployed agents, this is cost control rather than capacity improvement. A cascade of probes can control LOOP without rewriting the scheduler or changing the model's final response policy. The difficult dependency is access. Production providers rarely expose activation-level OCs. If vendors expose secure internal state signals, early fault detection could become part of the measurement layer for multi-step agents. Until then, the article offers a strong research result. Exit scraping is too late for many agent failures. Depteavikavé attacks the same long context service wall from a different angle than FREDEPIC. When the context grows beyond hundreds of thousands of tokens, the cash key value can consume more memory than the model weights. Existing compression schemes often apply a uniform budget across tiers and tokens, which can corrupt critical recovery details. Depteavikavé instead shares state between neighboring transformer layers using low-rank channel bases. The method separates two types of information. Activations that barely change between layers are absorbed into a shared basis. Token-specific residues remain only where attentional behavior is sensitive. Tokens that carry broad semantic state can share more aggressively. Tokens that anchor lexical search, code references, or retrieval behavior maintain narrow waste slots. The adaptive part comes from targeting residual budget with a per-token attention sensitivity signal rather than a global compression KNOB. This matters for long-term coding agents. A multi-hour session can include project-wide context, tool traces, DIFs, shell output, and user intent across multiple rounds. Meeting this context is not just about model capacity. It's about keeping the KV Workset within the memory and GPU bandwidth budget. Adaptive sharing could increase the number of competing long sessions or reduce the cost of serving each one. Engineering issues are runtime issues. Can the factored state flow incrementally as new tokens arrive? Serving engines like VLLM, TensorTTLLM or similar can adopt cross-layer sharing without breaking paged tension. Does the compression ratio survive Binsert, tool-heavy traces, and code recovery rather than just clean recovery benchmarks? If the answer is yes, long-term customer service gains a practical path to efficiency without requiring a new base model. Halboat is the first project on the radar because it is a concrete fars local client, not just a topic of discussion. The repo features a provider-neutral AI chat desktop surface where model connectors, project context, and user-managed keys can live in a single application. The main mechanism is local ownership of section state plus provider routing. A team can connect Antropic, an OpenAI-compatible endpoint, or a self-hosted model server, and maintain the same interaction layer for project work. The integration angle is direct. Halboat could become the desktop shell around internal template gateways and private MCP servers if maintainers can get reliable tooling support. Cocoro is the second project because it changes the way of deployment for voice agents. The ecosystem of repos around Cocoro and its ONNX packaging gives developers a compact speech synthesis component that can run on CPU hardware. The core engine is high-fidelity TTS with small parameters and style latent prosody plus optimized ONNX execution. The integration angle is a local voice layer for agents who already handle speech recognition or text chat. Instead of sending each response to a hosted TTS service, an application can synthesize locally, reduce latency, lower recurring API cost, and preserve privacy for sensitive utterances. Danus is the third project to watch, even before a broad ecosystem forms around it, because it offers a reusable memory standard. The central idea is a fact graph shared between parallel reasoning workers. The main mechanism is assertion-level coordination. Lemmas, definitions, and intermediate results become graph nodes and edges with provenance. No shifts buried within a chat stream. The integration angle extends beyond mathematics. Search agents, compliance review systems, and code analysis agents can use the same pattern when they need many workers to coordinate around verifiable assertions rather than loose summaries. Aeon 3 Mini is the model selected to accompany. It is available through the Open Router chat interface, and the DeepSec family multilingual behavior, and brings a 3-1K token context window for role-playing and storytelling workflows. The selected angle is not raw dominance in benchmarks, it is specialization. The model is shaped for character voice, continuity, and long interactive sessions. Whether you're building NPC dialogue, tabletop game wizards, interactive fiction, or persistent character chat, Aeon 3 Mini offers a hosted endpoint where narrative structure is part of the design rather than an afterthought. The local spotlight goes to Cocoro because it makes it easy to package a complete voice loop on consumer hardware. The practical angle of experimenting now is simple. Connect Kokoro with a TTS endpoint behind an existing on-premises agent, maintain CPU-intensive speech synthesis, and compare latency and audio quality with a hosted speech provider. The small footprint means it can run alongside a local chat model, a recovery service, or a desktop automation agent without turning the machine into a GPU-only appliance. For privacy-sensitive assistants, Kokoro is the kind of component that transforms local fast from a slogan into a deliverable mode of interaction. Fredepke and Depteav KV form the first line of additional research. Both attack the growth of KVCache, but emphasize different sharing structures. Frequency-driven depth sharing in one case, low-rank traversal bases with adaptive token residuals in the other. The combined message is that long context is now as much a service systems problem as it is a model training problem. If cache compression preserves critical recovery tokens, agent providers can serve longer sessions and more concurrent users from the same hardware envelope. Vaora is the second research candidate because reward design for language vision agents is moving from combined scores to separable terms. Separate visual grounding from outcome prediction Teams action Cleaner fault attribution This is important for browser control, robotics, remote desktop automation, and any VLM workflow where the model must both see the current state and predict what your next action will change. The early failure PROBS cascade is the third candidate because it moves cost control into the loop. Instead of waiting for a bad final response or visible tool failure, PROB reads internal representations and marks doomed trajectories early. The limitation is the provider's access to activation signals, but the research direction is clear. Agent runtimes need a way to stop expensive failures before they become long, polished failures. Codex Rust.143 makes remote plugins and system proxy routing default behaviors, while Hermes 7.7 and Cloud Code.197 improve traceability, sandbox control, and telemetry around agent work. Ion 3 mini gives narrative agents a long context model, focused on role playing through open router, with persona continuity as the key reason for deployment. Kokoro makes local speech synthesis practical for CPU fast voice agents who need lower latency, lower cost, and less cloud dependency. Antropic's Workspace Red paper gives interpretability teams a narrower target for reasoning probes and tool routing. Halboat Shows Demand for Local Fast, Provider Neutral Desktop Artificial Intelligence Customers with Brin-Oron-Key Routing Potential Future MCP. Fredepck and Depteav-KV point to KVCast's adaptive compression as a direct lever on long context service cost. Danus argues for fact graph memory when multi-agent systems need assertion provenance, not just message history. The stochastic model discovery paper pushes the evaluation of coding agents for repeated trials, variance bands, and factor effects. Rubente 1.0 tests whether coding agents can handle maintenance requests in native Russian on real repositories. Antropic's API migration friction shows why Frontier model upgrades require attention to prompt hierarchy and tool routing regression. Vaora separates visual grounding from outcome-action reasoning for VLM agents. The prompt injection leak on GitHub enforces per-call authorization for each agent tool invocation. Early failure probes show that wasted agent processing can be cut before the failure reaches the visible output stream. For links and deeper source context, look at the show notes at tobionfitnessstec.com. Thanks for listening to Edit Stack Daily. We'll be back soon.