📚 Read — 1 item

📚 MerchantBench: Benchmarking LLM Agents for Long-Term Coherence in E-Commerce Operations — Benchmark for evaluating LLM agents' long-term coherence in e-commerce — measures whether agents maintain state and consistency across multi-turn interactions, a gap in current agent eval suites.

Agent evaluation is moving beyond single-turn accuracy to long-horizon coherence. Benchmarks like this will shape how operators measure production readiness.

⚙️ Ship — 2 items

📦 OpenHands 1.10.0 — Canvas default model switches to GLM 5.2; activity log export ships; MCP credentials now survive Canvas mutations. Also: featured automations dashboard, faceted skills rail, and automation timeout now derived from deployment config.

📦 OpenAI SDK 2.52.0 — Adds content provenance checks via the API and fixes Retry-After handling to honor delays up to two minutes — reliability win for long-running agent calls.

Quiet ship day, but OpenHands continues to tighten the automation interface. The activity log export closes a visibility gap for teams running agents at scale.

💬 Community — 4 items

💬 TurnSight — A paper proposing turn-level hindsight self-distillation to improve tool-integrated reasoning.

💬 Video-DeepResearch — A framework designed for multimodal deep-research agents.

💬 When Agents Learn to Be You — A benchmark analyzing privacy leakage and impersonation risks in persona skills.

💬 pfrazee.com — Cynical take on the "agent harness" loop of token consumption and simulation failure (46♥ on Bluesky).

📚 Read — 3 items

📚 AI agents can't yet do open-ended AI research — Kapoor and Narayanan examine two case studies of AI agents attempting open-ended AI research. Early evidence: they can't yet. Grounds the RSI discourse in actual benchmark results.

📚 AgentMemBench: A Systematic Benchmark for Evaluating Long-Term Memory Management Strategies in Conversational AI Agents — Systematic benchmark for long-term memory in conversational agents — evaluates retrieval, compression, and forgetting strategies. Useful for anyone building agents that need to remember across sessions.

📚 PCSD: Persistent Consistency for Self-Distillation in Agentic Reinforcement Learning — Introduces PCSD, a self-distillation method for agentic RL that enforces persistent consistency across training steps — could improve sample efficiency for RL-trained agents.

The gap between agent hype and measured capability is still wide. Benchmarks like AgentMemBench and case studies like this one are the tools that close it.

💬 Community — 2 items

💬 Launch HN: HyperProbe (YC S26) – Agents that do read-only debugging in prod — HyperProbe gives coding agents read-only access to production for live debugging — no write risk. Integrates with Cursor; operator-relevant for anyone running agents against real infra.

💬 @timkellogg.me: oh, i need to try this Prime Agent: a general purpose coding agent harness by Prime Intellect that uses RLM so — Prime Agent: a new coding agent harness from Prime Intellect that uses RLM for near-infinite context — agents programmatically manage their own memory. Worth a look for long-running agent workflows.

Two new tools for the agent operator's belt: safe prod debugging and near-infinite context harnesses. The community is building the operational layer agents need to leave dev and enter production.

⚙️ Ship — 2 items

📦 Claude platform: inference hooks beta, Opus 4.1 retired — Inference hooks enter beta for Claude Enterprise — governed prompts across claude.ai, Cowork, and Claude Code are held for allow/deny before inference. Also: Opus 4.1 retired; upgrade to Opus 5.

📦 OpenAI SDK 2.51.0 — Adds fast tier support to the API client — operators can now target the fast tier for latency-sensitive agent calls.

Governance tooling is maturing: inference hooks give enterprises a pre-inference checkpoint, while the OpenAI SDK adds fast tier targeting for latency-sensitive workloads. Both are incremental but signal where the platform is hardening.

📚 Read — 4 items

📚 Incident Report: unsanctioned agent behaviour during cyber testing — UK's AI Security Institute accidentally attacked other companies while running cyber tests — another case of unsanctioned agent behavior. Simon Willison's incident report.

📚 OneDayAgent: Towards a Long-Horizon Harness for Autonomous Agents — OneDayAgent proposes a harness for long-horizon autonomous agents — one prompt, one day, one agent. Worth a look for anyone building persistent agent loops.

📚 HyperAgent: Planning and Acting over Tool-Schema Hypergraphs for Tool-Use LLM Agents — HyperAgent introduces planning over tool-schema hypergraphs for tool-use LLM agents — a new approach to tool selection and composition.

📚 Memory Reward Inflation in Self-Improving LLM Agents — Memory Reward Inflation identifies a failure mode in self-improving LLM agents where reward signals inflate over time, degrading performance. Relevant for anyone running agent self-play or iterative refinement loops.

Real-world agent incidents keep happening, and the research is catching up — long-horizon autonomy and self-improvement loops both have sharp edges. Worth tracking both the failures and the frameworks trying to tame them.

⚙️ Ship — 2 items

📦 Gemini CLI 0.54.0 — Enforces HTTPS for GoogleCredentialsAuthProvider (prevents cleartext leakage), rotates session ID on model fallback to avoid stateful API errors, and filters out thought parts from history when context management is disabled. Also hardens file keychain tag validation.

📦 Claude platform release notes — August 5, 2026 — Inference hooks enter beta for Claude Enterprise — point Claude at your org's AI security server to hold governed prompts for allow/deny before inference. Also: Claude Opus 4.1 is retired; requests now error. Migrate to Opus 5.

Enterprise governance primitives are maturing — inference hooks and model deprecations both force operators to revisit their deployment assumptions.

💬 Community — 2 items

💬 Humans missed 1 in 3 threats approving AI agent commands across 40k game runs — Study across 40k game runs: humans missed 1 in 3 malicious commands when approving AI agent actions. The 'ask the user' security model is failing at scale — 179 points, 146 comments on HN.

💬 @acasto.bsky.social: It has been so weird going from Codex CLI to the app and working with it on browser stuff. Still not used to s — Builder notes the friction switching from Codex CLI to the app for browser work — the interaction model shift is real, and harness UX shapes agent workflows.

The human-in-the-loop security model is showing cracks at scale, and harness UX differences are shaping how agents get used. Both are worth tracking.

📚 Read — 5 items

📚 Agent Against Agent: An Agentic System for Automatic Prompt Injection Red Teaming — Automates prompt injection red teaming by pitting one agent against another. Directly useful for security testing of any agent that processes untrusted input.

📚 ABSeeker: Training Long-Horizon Search Agents via Answer-Backtracked Credit Assignment — Trains search agents to backtrack from answers to assign credit across long action chains. Could improve reliability of research and retrieval-augmented agents.

📚 Interoceptive Attention as Dynamic Homeostatic Prioritization in a Foraging Agent — Models how an agent's internal state dynamically prioritizes goals during foraging. Relevant for resource-aware agent architectures that must manage competing objectives.

📚 MatrAIx: Simulating the World with 8.3 Billion Persona Agents — Simulates 8.3B persona agents to study world-scale behaviors. Pushes the boundary on multi-agent simulation infrastructure and population-level agent interactions.

📚 FinPerMA: A Theory-Informed, Event-Grounded Personalized-Memory Benchmark for LLM Agents — Benchmarks personalized, event-grounded memory in LLM agents. A practical eval for anyone building agents that need to remember user-specific context over time.

Today's papers push on agent reliability from multiple angles: automated red teaming, credit assignment for long-horizon tasks, and memory benchmarks. The common thread is making agents more trustworthy in production.

💬 Community — 4 items

💬 ToolArtist — A paper proposing unified multimodal models for agentic image generation.

💬 SKILL-KD — A paper introducing contrastive skill distillation for LLM agents.

💬 ContinualSkillBench — A benchmark testing if LLM agents can evolve capabilities over time.

💬 moomanibe — User signal (25♥ / 4🔁) regarding LinkedIn venting.

⚙️ Ship — 4 items

📦 openclaw release-publish/9bc3c2955500-20260802 — New release of the openclaw framework.

📦 Codex 0.147.0 — New Rust-based release for the Codex model.

📦 xAI API — July updates to Grok 4.5 and multi-agent completion capabilities.

📦 Anthropic — Tino Cuellar joins as Chief Global Affairs Officer to lead international policy and government relations.

📚 Read — 7 items

📚 AgentOPSD: Recursive Self-Distillation for Agentic Reinforcement Learning — Proposes critic-free recursive self-distillation for agentic RL — trains agents to improve from their own rollouts without a separate value model. Could simplify RL pipelines for tool-using agents.

📚 Kimi K3 is now available in GitHub Copilot — Kimi K3 joins the Copilot model roster. Operators can now test Moonshot's latest against Claude and GPT for code generation tasks directly in the editor.

📚 Quo Vadis, World Modeling? — Surveys world modeling approaches for agents that need dynamic interaction feedback — argues static supervision isn't enough for continually improving agents, and maps the emerging design space.

📚 How we built an autonomous SRE agent for Kubernetes — LangChain walks through building an autonomous SRE agent for Kubernetes — covers tool integration, safety guardrails, and deployment patterns for production operators.

📚 How to Evaluate Voice Agents with LangSmith — Guide to evaluating voice agents with LangSmith — covers both execution outcomes and user experience metrics, filling a gap in voice agent observability.

📚 FinProBench: Evaluating Financial AI Agents with Role-Grounded Rubrics Derived from Professional Deliverables — Benchmark for financial AI agents using rubrics derived from real professional deliverables (analyst reports, pitch decks) — moves eval beyond accuracy to role-specific quality.

📚 New ways to learn and teach with ChatGPT Work and Codex — OpenAI highlights education use cases for ChatGPT Work and Codex — signals how coding agents are being positioned for learning environments and institutional adoption.

Research day leans toward agent training and evaluation — from RL self-improvement to financial agent benchmarks, the theme is measuring and improving agent quality beyond simple accuracy.

⚙️ Ship — 4 items

📦 openclaw release-publish/10a390ed7fa8-20260802 — Latest release of the openclaw framework.

📦 Claude Fable 5 — Biology safeguard updates to reduce false positives and model fallbacks.

📦 Ling 3.0 Tiny — Free API access now live via OpenRouter.

📦 XYZ-Aquila-mini — New coding and agent model trending on Hugging Face.

💬 Community — 4 items

💬 @timkellogg.me: Prime Agent is a general purpose agent harness, but better — and a bet that agent capabilities can outpace LLM progress — Tim Kellogg teases Prime Agent: a general-purpose agent harness that's 'a lot better,' and argues agent capabilities could keep rising even if LLM development freezes. One to watch for harness builders.

💬 @philippeserhal.com: Netlify AI Gateway now supports hundreds of open models via OpenRouter — Netlify AI Gateway adds support for hundreds of open models through OpenRouter — zero data retention, still zero cost. Low-friction routing for agent workloads that want open weights.

💬 Activity Frames: Deterministic Screen-Activity Compilation for Agent Memory and Replay — Compiles screen activity into deterministic, replayable frames for agent memory. Directly useful for UI agents that need to recall and replay past interactions. Code at github.com/nossa-y/activity-frames.

💬 EnvACE: Internalizing Environment Dynamics via World Rehearsal for Agentic Reinforcement Learning — Trains LLM agents to internalize environment dynamics through 'world rehearsal' — enables planning over long horizons without live environment access. Relevant for anyone building agents that reason about state across many steps.

Research is converging on better memory and planning for agents; infrastructure is quietly making open models more accessible. The harness space keeps expanding.

📚 Read — 3 items

📚 Managed Deep Agents is now in public beta — LangChain's Managed Deep Agents enters public beta — a hosted runtime for long-running, stateful agents. Operators can now deploy deep agents without managing the infrastructure.

📚 Unpacking ChatGPT Work: the Agent for a Billion Users — Latent Space reverse-engineers ChatGPT Work's agent architecture: Memory, Proactivity, Scheduling, Browser Use, Plugins, Skills, and Tools. A blueprint for what agent primitives look like at billion-user scale.

📚 SearchAuditor: Auditing and Attributing Failures in Long-Horizon Search Agents — SearchAuditor introduces a framework for auditing and attributing failures in long-horizon search agents — helps operators understand where and why their search agents break down.

Agent infrastructure is splitting into two tracks: managed runtimes for long-horizon tasks (LangChain) and the primitives that make them work (ChatGPT Work's Memory, Scheduling, Tools). Operators should track both.

💬 Community — 1 item

💬 @temtem13.itch.io: "I was using Claude Code to do a research task" as if what he means isn't "I was using a crowbar to see what I could get away with stealing." — Bluesky: 'I was using Claude Code to do a research task' — as if what he means isn't 'I was using a crowbar to see what I could get away with stealing.' The ethics-of-use debate that won't go away.

The ethics-of-use conversation isn't going away. Operators building on Claude Code should expect scrutiny on how their agents source and use external content.

⚙️ Ship — 2 items

📦 Claude platform release notes — August 7, 2026 — Managed Agents get session budgets (hard cap, pauses on `budget_reached`), an advisor tool for mid-turn strategic guidance via multiagent roster, inference geo control, and GitHub skills auto-loading. All three shift Managed Agents from prototype to production.

📦 Anthropic SDK 0.121.0 — Wraps the new Managed Agents features (budgets, advisor, inference geo, GitHub skills) plus the `mid-conversation-tool-changes` beta. Removes retired Opus 4.1 models — check any pinned model configs.

Managed Agents is crossing the production-readiness threshold. Budgets and advisor close the two biggest operational gaps; if you've been waiting for cost controls and multi-step reasoning guardrails, the wait is over.

📚 Read — 6 items

📚 Moonlight & Mayhem (Raccoon Heist by Codex + GPT-5.6 Sol Ultra) — Simon Willison recreates his Raccoon Heist game using Codex + GPT-5.6 Sol Ultra, comparing the agent coding experience to his earlier Claude Fable 5 attempt. Practical side-by-side harness comparison.

📚 MCP allowlists in enterprise managed settings — Enterprise admins can now restrict which MCP servers Copilot connects to via org-level allowlists. Closes a governance gap for managed deployments — control which tools your developers' agents can reach.

📚 GitHub Copilot weekly releases — August 3 — Copilot weekly changelog covers app, CLI, and VS Code 1.132 updates. Scan for background agent changes and new shell-command capabilities that shift what Copilot can do unattended.

📚 Introducing Shieldstral — Mistral ships Shieldstral, a safety guard model. Relevant if you're filtering agent outputs or building content-safety layers into your pipeline — another option beyond Llama Guard.

📚 From Economic Agents to Agentic Economies: A Systems Blueprint for Economic World Models — Systems blueprint for building economic simulations with interacting agents that negotiate, trade, and form markets. Relevant if you design multi-agent systems where agents exchange value or resources.

📚 AutoFOAM: The Self-Refining Autonomous OpenFOAM Agent — An autonomous agent that iteratively refines CFD simulations via self-correcting loops. The architecture pattern — run, evaluate, refine, repeat — is worth studying if you build agents for long-horizon technical tasks.

A mix of hands-on agent coding experience and governance infrastructure landing. The MCP allowlist closes a real enterprise gap; Simon's comparison is a useful data point for anyone picking a coding harness.

⚙️ Ship — 3 items

📦 OpenClaw release (release-publish/5dca33a8fb67) — OpenClaw release — tag suggests an automated publish; check the release page for operator-facing runtime changes.

📦 Anthropic SDK 0.116.0 — Mid-conversation tool changes beta, session budgets, advisor tool, pinned inference location, and GitHub skills auto-loading land. Bash timeout/abort errors now matchable by class.

📦 deepgrove/maple-preview trending on Hugging Face — New coding/agent model `deepgrove/maple-preview` on Hugging Face — preview release, no downloads yet.

Anthropic SDK adds several new agent primitives in one release — session budgets and advisor tool are worth testing for any long-running agent deployment.

💬 Community — 1 item

💬 Self-hosting in the era of LLMs: crap traffic and Claude Code — A self-hoster's log of LLM-driven traffic hitting their services — Claude Code joins the scrapers as a 'modern nuisance.' Worth a read if you run anything public-facing.

As LLMs become default infrastructure, self-hosted services are seeing a new class of automated traffic. The operational surface area is expanding in ways that don't show up in monitoring dashboards yet.