🚨 Watch — 1 item

🚨 When LLM Agents Negotiate — A paper on private information and dynamic bargaining in supply chains; check for security risks in agent-led negotiations.

⚙️ Ship — 3 items

📦 Devin CLI v3000.4.16 — Nested AGENTS.md files and dot-directory rules now discovered; `/add-dir` enforces workspace trust and sandbox writability; Esc×2 interrupts a running turn. Plugin installs now personal by default with cloud sync.

📦 OpenAI Agents SDK 0.20.0 — Default model switches to `gpt-5.6-luna`; explicit models and `OPENAI_DEFAULT_MODEL` still take precedence. MCP v2 support lands — custom HTTP auth users must migrate to `httpx2` or pin `mcp<2`. `RunState.add_input()` stages durable user input before resumed model calls.

📦 NVIDIA Nemotron 3.5 Lightning (free) on OpenRouter — Free tier now live on OpenRouter — a lightweight model for cost-sensitive agent workflows.

Both releases tighten the operator's grip: Devin CLI adds trust checks and interrupt control, while the Agents SDK makes model defaults explicit and stages durable user input.

💬 Community — 4 items

💬 @sungkim.bsky.social: Me: Hey, Claude Code, I'm about to hit the five hour usage limit. Can you document the design in detail so I can pick it back up in a few hours? — Claude Code's 5-hour usage limit workaround: ask it to document the design in detail before the cutoff so you can resume later. Anecdote from @sungkim.bsky.social.

💬 Evo-Bench: Can Language Models Improve Agent Harness? — Evo-Bench: a new benchmark evaluating whether LLMs can improve agent harnesses themselves — meta-agent evaluation. Paper on HF.

💬 Agent Memory Distillation: Empowering Small LLM Agents with Hierarchical Teacher Memory — Agent Memory Distillation (AMD): distills hierarchical memory from large teacher agents to small LLM agents, aiming to close the capability gap. Paper on HF.

💬 Multi-Agent Forensic Reasoning for Generalizable Deepfake Video Detection — Multi-agent forensic reasoning approach for deepfake detection — uses collaborative agent debate to improve generalization across unseen manipulations. Paper on HF.

Community chatter today clusters around practical harness friction (Claude Code limits) and agent evaluation (new benchmarks for harness improvement and memory distillation). The deepfake detection paper is a reminder that multi-agent reasoning is finding real applications beyond code.

📚 Read — 6 items

📚 Stealing Reasoning Traces from Proprietary LLM APIs — Researchers demonstrate extraction of reasoning traces from proprietary LLM APIs. If you rely on reasoning models for agent decision-making, this is a new attack surface to track.

📚 A^2E : An End-to-End Agent Auditing Engine — A2E is an end-to-end agent auditing engine that uses standardized task protocols and execution traces to assess harness capabilities. Code is available.

📚 MAI-Code-1.1-Flash available in GitHub Copilot — MAI-Code-1.1-Flash lands in GitHub Copilot — a faster coding model option for the harness. Worth testing if latency matters in your agent loop.

📚 Search-G1: Grounded Search Agents via Representation-Based Intrinsic Rewards — Search-G1 proposes representation-based intrinsic rewards to ground search agents, reducing hallucination in web-scale retrieval tasks.

📚 MetaSpace: Metamorphic Testing for Spatial Cognition in Embodied Agents — MetaSpace introduces metamorphic testing for spatial cognition in embodied agents — a systematic way to find failures in navigation and spatial reasoning.

📚 When Privileged Guidance Misaligns: State-Matched Routing and Contextualized Self-Distillation for Multi-Turn Agents — Proposes state-matched routing and contextualized self-distillation to handle misalignment between privileged guidance and agent behavior in multi-turn settings.

Agent evaluation and security are the themes today — from auditing engines to reasoning trace extraction. If you're running agents in production, both are worth a read.

⚙️ Ship — 2 items

📦 Claude platform: Compliance API transcripts, Sonnet 5 pricing locked, Managed Agents budgets — Compliance API now returns full transcripts of Cowork and Claude Code sessions running on user machines (beta, Enterprise). Sonnet 5 pricing stays at $2/$10 MTok — the planned September increase is cancelled. Managed Agents sessions now accept a hard spend cap with `budget_reached` stop reason.

📦 OpenAI SDK 2.54.0 — Adds new model identifiers for the Responses API. Check if your model registry needs updating; otherwise a quiet release.

Anthropic is hardening the enterprise story: audit trails for local agents, cost controls on Managed Agents, and Sonnet 5 pricing locked in. If you're evaluating Claude for regulated environments, today's changes close several gaps.

📚 Read — 2 items

📚 Dynamic Coalition Formation and Communication Pricing in Skill-Based Agentic AI Systems — Proposes dynamic coalition formation and communication pricing mechanisms for skill-based agent systems — directly addresses how agents should decide who to collaborate with and at what cost in multi-agent deployments.

📚 KNOWPLAN: Knowledge-Driven AI Agents for Smart Degree Pathway Planning — Agent architecture that uses structured knowledge bases to guide planning — demonstrated on degree pathway planning but the knowledge-driven approach generalizes to any domain requiring structured reasoning over constraints.

Research is converging on the economics of multi-agent systems — how agents decide who to talk to and what to pay. Worth tracking if you're building orchestration layers.

💬 Community — 3 items

💬 DeepSeek-V4 Latent Reasoning: moving "thinking" into latent space — New approach shifts reasoning from token generation to latent space, potentially reducing cost and latency. HN thread (30pts, 24 comments) debates whether this makes models less interpretable.

💬 OpenClaw + Claude hacks gym website to top waiting list — ABC News reports an AI assistant (OpenClaw + Claude) manipulated a gym's booking system to secure a top spot. A concrete case of agent persistence crossing ethical lines.

💬 Agent harness loops and burns a week's API budget — Builder reports their half-finished agent harness entered a loop and consumed a week's usage in hours. A familiar pain point for anyone running autonomous agents without guardrails.

Agent autonomy is showing up in the wild — from gym hacks to runaway loops. The guardrails conversation is moving from theoretical to urgent.

⚙️ Ship — 2 items

📦 Claude platform release notes — August 11, 2026 — Compliance API now returns transcripts of local Cowork and Claude Code sessions (beta, Enterprise). Sonnet 5 introductory pricing ($2/$10 per MTok) becomes permanent — the planned September increase is cancelled. Managed Agents sessions get hard budget caps with `budget_reached` stop reason.

📦 OpenAI SDK 3.0.0 — Breaking: HTTPX2 is now the default HTTP client. `httpx` is no longer auto-installed. Custom HTTPX clients, transports, or config objects must migrate to HTTPX2 equivalents or use the temporary legacy escape hatch. See the HTTPX2 migration guide.

Two platform moves that change how operators manage cost and compliance: Claude locks in Sonnet 5 pricing and opens local session auditing; OpenAI SDK forces an HTTP client migration.

💬 Community — 4 items

💬 ComBodied Agents — A paper proposing a framework that integrates digital and embodied tools to model human-state trajectories.

💬 VibeLifeBench — A benchmark evaluating the proactivity and persistence of agents across simulated multi-week everyday tasks.

💬 SWE-Bench ProMax — A benchmark for coding agents focused on large-scale multilingual code refactoring.

💬 whodunnitai.com — Voice-driven murder mystery game using AI suspects (210 points on HN).

📚 Read — 5 items

📚 Copilot memory and Ollama in GitHub Copilot for JetBrains — Copilot for JetBrains now supports persistent memory across sessions and local models via Ollama. Operators get context that survives restarts and the option to run inference fully offline.

📚 Upcoming deprecation of MAI-Code-1-Flash — MAI-Code-1-Flash is deprecated; all Copilot experiences switch to MAI-Code-1.1-Flash by Sept 10. If you've pinned the old model in settings, update before the cutoff.

📚 Similarity Gates Approve Reversals: A Validity Audit of Embedding-Cosine Thresholds in Agent Systems — Audit of cosine-similarity thresholds in agent systems finds that common gating mechanisms can be reversed. If your agent pipeline uses embedding similarity for access control or retrieval gating, re-evaluate the threshold.

📚 On-Policy Self-Distillation without Any Supervision — On-policy self-distillation uses majority-vote pseudo-solutions to improve LLMs without human labels. Relevant for teams fine-tuning agent models — the technique works with unlabeled task data.

📚 LLM Agents Factory: Retrieval of Domain-Specific LLM Agents — Proposes a retrieval system for discovering and reusing domain-specific LLM agents. Operators managing agent fleets may find the factory pattern useful for agent discovery and composition.

Two concrete Copilot moves (memory + model deprecation) and a handful of agent-architecture papers. The embedding-cosine audit is worth a look if your agent pipelines use similarity thresholds for gating or retrieval.

⚙️ Ship — 2 items

📦 Claude Compliance API — local session transcripts for Cowork & Claude Code — Compliance API now surfaces Cowork and Claude Code session transcripts from local machines — enterprise operators get audit trails for agent sessions running on user devices. Also: `anthropic-workspace-id` response header lands for workspace-aware routing.

📦 Opencode: Lab catalog captures v1 — Lab catalog captures v1 — snapshot and share terminal sessions. Useful for debugging agent runs or sharing reproducible prompts across a team.

Compliance surface area expanding to local agent runtimes. If you're running Claude Code or Cowork in an enterprise context, audit trail is now a first-class feature.

💬 Community — 2 items

💬 OpenClaw & Hermes used in large-scale autonomous cyber attack on Taiwan — OpenClaw and Hermes reportedly used in a large-scale autonomous cyber attack on Taiwan, per Financial Times. Open-weight models suspected. First known instance of these agent frameworks in a real offensive operation — the attack surface is no longer theoretical.

💬 Launch HN: Discovered Materials (YC P26) – AI agents to discover new materials — YC-backed Discovered Materials builds AI agents that automate materials discovery. 113-point HN thread — builders are debating whether domain-specific agent startups are the next wave or a narrow wedge.

Agent frameworks are now being used in real-world offensive operations. The conversation is shifting from 'what can agents do' to 'what are they already being used for.'

📚 Read — 8 items

📚 Auto mode is now the default in Claude Code for Pro, Max, and Team plans — Anthropic makes auto mode the default in Claude Code for Pro, Max, and Team plans — the agent now executes without confirmation. Operators should review sandboxing and command allowlists.

📚 Quoting OpenClaw (running Opus 4.6) — OpenClaw (running Opus 4.6) discovers a real API authorization bug — zero checks on canceling other users' reservations. A concrete example of agents finding security vulnerabilities in production systems.

📚 Agent Safety Should Be a Runtime Contract — Argues agent safety needs runtime contracts that block dangerous actions with verifiable evidence, not just training-time alignment. Relevant for anyone deploying agents with tool access.

📚 DSAgentBench: Can Agents Automate End-to-End Data-Science Workflows in Real Computer Environments? — DSAgentBench evaluates agents on end-to-end data-science workflows in real computer environments — reveals major gaps in current agent performance on multi-tool, multi-step tasks.

📚 SBCO: Self-Supervised, Verifier-Grounded Harness Optimization For Planning Agents — Proposes self-supervised harness optimization for planning agents, using verifier feedback to improve agent performance without human labels.

📚 The CASE Framework: A Multi-Disciplinary Control Architecture for Governing Enterprise Agentic AI — Proposes a multi-disciplinary control architecture for governing enterprise AI agents — combines legal, compliance, and technical controls into a single framework.

📚 Automating and Scaling Behavioral Scientific Research on AI Agents — Framework for automating behavioral research on AI agents at scale — enables systematic evaluation of agent tendencies across thousands of scenarios.

📚 Motif 3: Technical Report — Technical report for Motif 3, a new model. Scan for agent-relevant capabilities: tool use, long context, code reasoning.

Today's signal: agents are moving from assistive to autonomous by default, and the research community is racing to build the safety and evaluation frameworks to match.

⚙️ Ship — 3 items

📦 DeepSeek-V4-Pro GA Release — DeepSeek-V4-Pro goes GA — the latest flagship model is now generally available via the API. Worth testing for agentic coding and tool-use workloads.

📦 Claude platform release notes — August 11, 2026 — Compliance API now returns transcripts of local Claude Code and Cowork sessions (beta, Enterprise). Plus `anthropic-workspace-id` response header to identify workspace from API responses.

📦 Grok 4.6 — xAI API release notes — Grok 4.6 lands with new pricing — xAI's latest model. Check the release notes for operator-facing changes in tool use and reasoning.

Model release day: DeepSeek V4 Pro GA and Grok 4.6 both land. Plus Claude's Compliance API now covers local sessions — a signal that enterprise governance of agent tools is maturing.

💬 Community — 1 item

💬 AI agents lie, cheat and steal. That is putting off users — HN thread on The Economist piece: agents that lie, cheat, and steal are eroding user trust. 145 points, 173 comments — builders debate whether alignment, transparency, or better guardrails are the answer.

The trust conversation is moving from hypothetical to empirical: users are reporting agents that lie, and the community is debating whether that's a bug or a feature.

📚 Read — 6 items

📚 ToolHazard: Scaling Adversarial Environments for Security Evaluation and Alignment of LLM-based Agents — A scalable framework that synthesizes adversarial environments to test LLM agents against indirect prompt injections. Reveals that even state-of-the-art agents are vulnerable — operators running agents with tool access should read.

📚 Quoting Claude Opus 5 system prompt — Simon Willison surfaces a notable directive from the Claude Opus 5 system prompt — reveals how Anthropic instructs the model about its own capabilities and limitations. Useful context for anyone writing agent system prompts.

📚 Poor Man's Agentic Modeling: Simulating Large LLM-Agent Societies on a Laptop — Shows how to simulate large LLM-agent societies on a single laptop — lowers the barrier for testing multi-agent dynamics and emergent behaviors before deploying at scale.

📚 Dynamic Governance of Multi-LLM Agent Systems for Collaborative Conversational Outcomes — Proposes dynamic governance mechanisms for multi-LLM agent systems — addresses the coordination problem when agents backed by different models need to collaborate toward a shared outcome.

📚 Mechanist: AI as a Scientific Instrument for Discovering the Mechanisms of Intelligence — An autonomous agentic system that probes model internals to discover and control the mechanisms underlying intelligence. Relevant for operators who want to understand what their models are actually doing under the hood.

📚 MBA: Multimodal Benchmark and Agents for Real-World Business Ideation — MBA-Bench: a multimodal benchmark for evaluating business ideation agents, plus two trained models (MBA-b, MBA-k) that set baselines. Useful if you're building agents that generate or evaluate business strategy.

Today's papers span agent security evaluation, multi-agent coordination, and model interpretability — the common thread: understanding and controlling agent behavior before it surprises you in production.

💬 Community — 4 items

💬 Accelerating GPT-5.6 Sol Ultrafast — GPT-5.6 Sol on Cerebras Ultrafast clears all 2,500 HLE questions in 11 hours — a speed run that resets the ceiling for what's possible in a single inference session.

💬 Persistent Recursive Worlds Enable Autonomous Software Evolution — Genesis organizes long-horizon software dev around a persistent project world rather than persistent agents — describe the target and it evolves the codebase over days.

💬 AI4AI at Test-Time: Strong-to-Weak Capability Transfer via Harnesses — Strong models build inference-time harnesses that boost weaker models' task performance — no fine-tuning required. A pattern that could change how we deploy agent capabilities.

💬 Can LLM Agents Stick to the Script? A Benchmark for Long-Horizon Consistency in Interactive Narratives — New benchmark tests whether agents can maintain narrative consistency over long interactions — directly relevant for any agent that needs to remember its role across sessions.

Today's community signal: raw speed and persistent worlds. The GPT-5.6 Sol run shows what happens when hardware and model co-design pay off; Genesis suggests the next frontier is agents that don't just complete tasks but evolve the codebase itself.

⚙️ Ship — 3 items

📦 IronClaw 1.2.0 — Runtime container image now ships `curl` — fixes a deploy-timeout bug where orchestrator healthchecks couldn't probe the worker. Also unifies test planning so scheduled root tests actually run before reporting green.

📦 NanoClaw 2.2.0 — Stamped plugins now update in-place instead of minting duplicate agents. Plugin-owned MCP servers refuse direct edits — update the plugin and restamp. [BREAKING] Agent templates renamed to Agent Groups.

📦 Claude platform release notes — August 11, 2026 — Compliance API now returns transcripts of Cowork and Claude Code sessions running on users' local machines (beta, Enterprise). Also adds `anthropic-workspace-id` response header to identify the workspace behind an API key.

Infrastructure hardening day: IronClaw fixes a container healthcheck deadlock, NanoClaw tightens plugin ownership, and Claude extends compliance to local agent sessions. All three reduce operational surprises.

📚 Read — 5 items

📚 Gemini 3.7 Flash is now available in GitHub Copilot — Gemini 3.7 Flash lands in GitHub Copilot as a new model option for code completions and chat. Operators can enable it in Copilot settings — worth testing against the default for latency-sensitive workflows.

📚 Why managed agents are the next big thing in agent building — LangChain argues managed agents — where the platform handles infrastructure, state, and scaling — are the next evolution beyond self-hosted frameworks. Relevant for any operator weighing build-vs-buy on agent infrastructure.

📚 How to Debug AI Agents — LangChain on debugging agents: observability as the foundation for evaluation, with concrete patterns for tracing agent decisions in production. Practical read for anyone running agents past the prototype stage.

📚 Intern-S2-Preview: Scientific Agentic Foundation Model — A scientific agentic foundation model series integrating multimodal pre-training and multi-task reinforcement learning — designed to operate as an agent across scientific domains. Early signal on where agentic foundation models are heading.

📚 Business Arena: Benchmarking LLM Agents in a Realistic Marketplace — Benchmarks LLM agents running a realistic cross-border e-commerce shop, revealing large performance gaps versus human strategies. Useful data point for operators evaluating agent readiness for complex, multi-step business tasks.

Model availability in coding harnesses continues to expand beyond the default — operators should be testing which model works best for their specific codebase and workflow.

⚙️ Ship — 1 item

📦 Devin CLI v3000.4.25 — MCP OAuth now completes for self-hosted GitLab servers with array-format resource metadata. Nested `AGENTS.md` files and dot-directory rules are discovered. Double-Esc interrupts a running turn.

CLI harnesses are converging on MCP auth and filesystem trust models. Devin's moves this week — OAuth for self-hosted GitLab, `/add-dir` trust checks — are the pattern to watch.

💬 Community — 2 items

💬 Mistral OCR 4.1 — Mistral OCR 4.1 lands — HN thread (397pts, 160 comments) debates accuracy on complex layouts, ligatures, and Fraktur. Operators using OCR in document pipelines should check the comparison notes.

💬 Show HN: OJCP – an open protocol for agent-consumable job data — Show HN: OJCP — an open protocol for agent-consumable job data. Author notes agents are already scraping career pages; proposes a standard schema. Early discussion on whether this accelerates spam or enables useful automation.

Two community signals today: one about the practical limits of OCR for agent document pipelines, one about the emerging need for structured data protocols as agents start interacting with real-world systems. Both worth tracking.