🚨 Watch — 1 item

🚨 MCP Python SDK: Unbounded Session Accumulation in StreamableHTTP Transport Enables Resource Exhaustion — Unbounded session accumulation in the StreamableHTTP transport enables resource exhaustion against MCP Python SDK servers — CVE-2026-59951, GHSA-84m7-p3x7-pcfv, medium severity. Anyone running that transport should check the advisory for affected and fixed versions.

📚 Read — 1 item

📚 3DCodeVerse: Exploring 3D Worlds as Code — Generates 3D worlds as code only — no image, video, or 3D diffusion models — and ships datasets plus automated verification harnesses for evaluating the output.

⚙️ Ship — 2 items

📦 Amp: See Your Runner's Desktop — You can now see and control your macOS or Linux runner's desktop — useful for trying native apps Amp builds for you, like iOS apps. Turn on Share This Mac's Display under Settings › Runner in the macOS app, or start your runner with amp --no-tui --desktop; the Amp client connects directly to the runner.

📦 OMP 18.12.0 — Breaking: PI_CODEX_ZSTD is replaced by PI_REQUEST_COMPRESSION — set PI_REQUEST_COMPRESSION=0 to send Codex (and every other) request body uncompressed. Also fixes Cursor turns failing with PROVIDER_ERROR when a follow-up arrived while the previous turn was still streaming, and adds ThinkingContent.narration for Anthropic/Bedrock between-tool narration blocks.

💬 Community — 3 items

💬 DeepSeek search and fetch – $0.75 per 1000 successful requests — DeepSeek search and fetch at $0.75 per 1000 successful requests — 443 points and 214 comments on HN, a concrete pricing point for anyone wiring web search into an agent stack.

💬 Codex Desktop continuously generates Crashpad pending dumps — Codex Desktop continuously generates Crashpad pending dumps — 161 points and 80 comments on HN, where operators trace disk pressure from running agents to causes including their git worktree approach.

💬 Trying to get jev to drive Claude Code — Trying to get jev to drive Claude Code — 92 likes / 15 reposts on Bluesky.

Both HN threads are about what agents cost to run: $0.75 per 1000 successful requests for DeepSeek's search and fetch on the API side, and disk pressure from Codex Desktop's Crashpad dumps and agent git worktrees on the local side.

📚 Read — 1 item

📚 GroundGUI: GUI Grounding by High-Quality Data and Distillation-to-Reinforcement Learning — GroundGUI pairs the ScreenRef data with a distillation-to-reinforcement-learning pipeline to train GUI grounding models — the instruction-to-screen-element mapping GUI agents run on.

💬 Community — 3 items

💬 Llama.cpp on a jailbroken PS5: run any GGUF model natively on the console — HN thread at 372 points / 85 comments on running any GGUF model natively on a jailbroken PS5; the top comment argues the project's primary site should not be GitHub.

💬 Designing more efficient quantum circuits with Claude — Anthropic's post drew 172 points / 37 comments on HN; the thread notes it builds on Google's Kickmix, an assembly language for classically simulatable quantum circuits.

💬 cuck coding: 'you just sit in a chair and watch, no touching' — Bluesky meme (70♥ / 15🔁) on the sit-back-and-watch school of coding with agents: 'you just sit in a chair and watch, no touching.'

⚙️ Ship — 2 items

📦 Pi 1.3.0 — Prompt-cache fixes lead: changing the thinking level mid-session on OpenAI GPT-6 (API key) now keeps the cache, and tools added mid-conversation on Kimi Code no longer invalidate it — but models.json base URL overrides must now include /v1. Also new: sessionShareDestination picks the /share upload target (auto, gist, or radius), model switching drops the previous model's thinking instead of showing it to the new model as text, and resumed sessions keep their active tools with --tools +name/-name stacking on top.

📦 Anthropic prepares to release Claude Fable 5.5 next week — Claude Fable 5.5 has reportedly reached Anthropic's partners ahead of a release next week; six early access codenames spotted in Devin CLI, plus one in a GitHub model picker, suggest it may not be the only model on the way.

📚 Read — 2 items

📚 CLASHarness: Benchmarking Multi-Agent Conflict in Computer Use via a Controllable and Verifiable Generation Pipeline — Benchmark paper introducing a controllable, verifiable pipeline for generating multi-agent conflict scenarios in computer-use tasks — the read for anyone orchestrating more than one computer-use agent.

📚 OmniVerifier-C1: Decomposed Graded Verifier for Rule-Governed Generative World Modeling — A graded visual verifier that decomposes complex multimodal rules into pointwise scores and uses decoupled reinforcement learning to improve rule-governed generative world modeling.

Both papers are verification work: CLASHarness generates multi-agent conflict scenarios in computer use via a controllable, verifiable pipeline, while OmniVerifier-C1 decomposes rule-following in generative world models into pointwise graded scores.

⚙️ Ship — 2 items

📦 OMP 18.11.0 — Breaking: ApprovalDetailsOptions removed — formatApprovalDetails takes only args again. Also fixes OpenAI/Codex remote compaction refusing or truncating sessions that fit the window: preflight counted the JSON-escaped request, inflating quote-, backslash- and newline-heavy tool output such as disassembly.

📦 Pi 1.2.0 — Per-key OpenRouter model lists (only the models that key may use, with context size, max output, and prices), $schema validation for settings.json, models.json, keybindings, and themes, and PI_CLIPBOARD_OSC52=1 to send copied text to a remote client's clipboard.

Both harness releases extend their extension surfaces: OMP 18.11.0 adds AgentTool.delegationGuidance for same-name extension replacements that delegate to the native tool, and Pi 1.2.0 adds before_agent_start for extension-triggered runs.

🚨 Watch — 2 items

🚨 Degraded availability for GitHub Actions, GitHub Pages and Webhooks — GitHub Actions, Pages, and Webhooks hit by degraded availability — CI runs and webhook-triggered deploys may be affected. Incident still open at time of posting.

🚨 Cursor: elevated errors for Anthropic Opus 5.5 — resolved — Elevated errors for Anthropic Opus 5.5 hit Cursor CLI, Cloud Agents, and cursor.com; fix implemented 13:35 UTC, incident resolved 13:52 UTC.

💬 Community — 1 item

💬 If you don't run your own LLM proxy, how do you know whats going on? — HN thread, 281 points and 199 comments. The operator question in the thread: for agents running behind LiteLLM or similar, are you logging full request and response bodies or just metadata?

📚 Read — 3 items

📚 MAI-Gym: An Open All-in-One Platform for GUI Agent Evaluation and RL Training across Mobile and Computer Environments — Open platform unifying GUI agent evaluation, data collection, and online RL training across mobile and computer environments.

📚 Tree-Aware Posterior Distillation for Long-Horizon Agentic Reinforcement Learning — TreeOPSD: tree-structured self-distillation for long-horizon agentic RL that trains without a critic model.

📚 Stack Overflow for Agents triples signups with ChatGPT plugin — OpenAI customer story: Stack Overflow for Agents tripled signups after shipping a ChatGPT plugin.

Both papers target RL for agents — MAI-Gym supplies the open GUI eval and training environment, TreeOPSD the long-horizon training method that drops the critic.

💬 Community — 1 item

💬 Websites are closing their doors to AI agents — HN discussion of The Verge's report — 153 points, 149 comments; the top comment calls out a 'you can't scrape us but we can scrape you' mentality and small businesses whose sites 'get absolutely hammered'.

⚙️ Ship — 3 items

📦 Amp is now an agent in Linear — Assigning an issue to Amp starts a new Agent Session in Linear, mentioning Amp kicks one off, and the session appears in your Amp sidebar, fully viewable in Amp.

📦 Kiro 1.4.0 — Memory Across Chat Sessions — Brings Memory to every Kiro IDE user; also makes find more reliable in long Markdown Review documents and fixes Kiro failing to start after a .deb install on Ubuntu 24.04.

📦 Copilot model policy API for enterprise, orgs, and teams — You can now use the REST API to retrieve and update Copilot model policies for enterprises, orgs, and teams — model policy management is scriptable.

📚 Read — 4 items

📚 Self-Evolving Coding Agents via Harness-Policy Optimization — Self-HPO lets the coding agent improve its own harness, then trains the policy — a recursive framework that alternates harness and policy optimization for coding agents.

📚 τ-Emotion-Bench: Exposing the Interaction Fragility of LLM Agents through Real-World User Emotions — Benchmark exposing LLM agents' interaction fragility in long-horizon tool-calling tasks under real-world user emotions.

📚 Librarian Agents: Training LLMs to Curate Long-Context Workspaces that Anticipate Future Queries — Trains builder agents with reinforcement learning to reorganize document corpora into reusable hierarchical workspaces that anticipate future queries.

📚 New paper: Emerging trends in AI agent evaluation — Chen, Rabanser, Kapoor, and Narayanan map the methods, questions, and ecosystem of AI agent evaluation — amid growing calls for independent eval, what constitutes a 'good' one remains unclear.

Two of today's papers train agents to reshape their own working environment — Self-HPO has the coding agent improve its own harness, Librarian Agents reorganizes document corpora into workspaces that anticipate future queries.