⚙️ Ship — 3 items

📦 Codex 0.148.0 — Hooks go async and can call MCP tools; `/export` saves TUI sessions to Markdown, `/status` shows thread cost, and Amazon Bedrock Runtime joins as a built-in provider.

📦 Claude platform release notes — August 18, 2026 — Workbench is now Playground in the Claude Console — supports every Messages API parameter, shows the full SDK request and API response per run, and includes templates for code execution and web search.

📦 OpenAI SDK 3.3.0 — OpenAI Python SDK 3.3.0 adds named data-residency endpoints, letting orgs pin API traffic to specific regions for compliance.

Between Codex's async hooks and OpenAI's data-residency endpoints, operators get more control over how agent traffic runs and where it lands. Worth testing before your next pipeline refactor.

📚 Read — 4 items

📚 Agentic ESOpt: Fine-Tuning Long-Horizon LLM Agents with Minimal GPU Requirements — Agentic ESOpt fine-tunes long-horizon LLM agents with evolution strategies and trajectory-level rewards — full-parameter tuning without the GPU bill. Worth a look if you're training agents on limited hardware.

📚 From Sequence to Structure: Relational Uncertainty Propagation for LLM Agents — RUPA models agent trajectories as relational graphs to propagate uncertainty across long runs — better failure detection and confidence calibration for multi-step agents.

📚 ASI-Bench: At the Dawn of Artificial Superintelligence — ASI-Bench evaluates agents on scientific exploration and innovation tasks — a benchmark for measuring how far current agents are from autonomous research.

📚 AutoMem: A Text-Gradient Recursive Self-Improvement Framework for Automated Memory Architectures Search — AutoMem uses text-gradient recursive self-improvement to automatically search memory architectures for agents — memory design as an optimization problem rather than hand-tuning.

Today's Read lane is about making agents more reliable and cheaper to train: uncertainty propagation, automated memory search, and low-GPU fine-tuning all point the same direction.

💬 Community — 1 item

💬 Devin cuts GPT-5.6 Sol pricing 70% — Devin drops GPT-5.6 Sol pricing 70% — worth checking if your Devin workloads are on that model; HN thread is light on comments so far.

Pricing pressure on coding agents is real; when a major harness cuts a model's cost 70%, it's worth re-evaluating your own per-task economics.

⚙️ Ship — 2 items

📦 OpenAI Agents SDK 0.22.0 — OpenAI Agents SDK 0.22.0 — guardrail-rejected tool output is now redacted from replay/persisted state, and non-streaming Responses with terminal `failed`/`incomplete` raise `ModelBehaviorError`. Breaking: if you pass an explicit `openai_client` with `organization`/`project`, move those to the `AsyncOpenAI` client.

📦 Anthropic SDK 0.118.0 — Anthropic SDK 0.118.0 — adds files/memory store API surface and workspace-ID helpers; session-runner retries tool-result sends for the lease TTL, and docs now warn that blocking tool bodies stall the worker heartbeat.

Runtime-hardening day: both SDKs tighten state handling and provider contracts — worth a quick migration check if you pass explicit clients or rely on replayable traces.

💬 Community — 3 items

💬 LEGO-RL: Harness-Native Reinforcement Learning for Coding Agents — Open-source framework for harness-native RL on coding agents — in-process LLM proxying and sandbox orchestration let you policy-gradient-train the agent inside its actual harness, not a stripped-down eval loop.

💬 Agent Lightning v1.0: Towards Harnessed Agentic RL — Reproducible RL for arbitrary agent harnesses — trains the full agent stack (harness + LLM) and reports substantial coding-agent gains. Worth a look if you've been RL-tuning models in isolation.

💬 StartupBench: Benchmarking General-Purpose Agents on Market-Validated End-to-End Workflows — Benchmarks general-purpose agents on market-validated startup workflows end-to-end; even top models finish only ~30% of tasks. A reality check on how far agents are from replacing a junior operator.

The RL-for-agents field is converging on harness-native training. If you're building coding agents, the question is no longer whether to RL-tune, but which harness integration you'll adopt.

📚 Read — 4 items

📚 HarnessRisk: A Lifecycle-Oriented Benchmark for Agent Harness Safety — Benchmarks agent harness safety across six operational phases — configuration vulnerabilities and detection gaps are the weak points, not model behavior.

📚 MCP in Orbs — Amp adds remote MCP server support across orbs, the TUI, and Puck — Streamable HTTP with OAuth. Connect hosted servers from ampcode.com/settings/mcp-servers.

📚 Replit expands access to software creation with GPT-5.6 Luna — Replit adds GPT-5.6 Luna as a model option for software creation — a new capability for Replit Agent users.

📚 Cross-Model Memory Transfer via Target-Side Reader Adaptation — Cross-model memory transfer via target-side reader adaptation — lets a model inherit memory from another without retraining, relevant for agent memory portability.

Harness safety and MCP plumbing are the day's themes — operators should audit config surfaces and expect MCP support as default.

⚙️ Ship — 4 items

📦 openclaw release-publish/abd63c06c2f5-1786646045 — New release.

📦 Codex 0.147.0 — New Codex release.

📦 Gemini CLI 0.56.0 — New Gemini release.

📦 Claude Platform — Files API and Enterprise Admin API endpoints are now generally available.

💬 Community — 3 items

💬 SkillForge: Self-Distilling Agents for Project-Specific Issue Resolution — SkillForge lets agents self-distill repo-specific skills by synthesizing issues, then reusing them to resolve real ones — a step toward agents that onboard themselves to a codebase.

💬 Demystifying Agent Skills: Why They Work-Until They Don't — Skills work by anchoring execution procedurally, not by adding knowledge — and they break when the procedure doesn't match the task. Useful for deciding when to encode a skill vs. rely on the model.

💬 Harness the Memory: A Holistic Evaluation of Memory Substrates in Memory Agents — Controlled eval of memory substrates for long-horizon agents shows no single substrate wins — trade-offs depend on task regime, pushing toward adaptive substrate selection.

Today's papers converge on a theme: agents get better by learning their environment (repo skills, memory substrates), but the gains are regime-dependent — test before you standardize.

📚 Read — 5 items

📚 Towards Safer RAG: Only Agents Capable of System 2 Thinking may Access Untrusted Documents — Proposes gating untrusted document access to agents capable of System 2 reasoning, arguing fast pattern-matching agents are vulnerable to prompt injection in RAG pipelines.

📚 KernelArc: A Multi-Agent Framework for GPU Kernel Optimization — Multi-agent framework where specialized agents collaborate on GPU kernel optimization; reports speedups over single-agent baselines and existing autotuners.

📚 SkillEffect: Checked Lowering for Memory-Bounded Agent Tools — Introduces checked lowering for agent tools, enforcing memory bounds at compile time to prevent unbounded memory growth in long-running agents.

📚 GxP-Agent: Process-DAG Topology for Reliable Clinical Trial Programming with LLM Agents — Structures LLM agents with process-DAG topology for clinical trial programming, targeting reliability in regulated GxP workflows.

📚 aDSL: Agentic 3D Creation via Joint Agent-Program Design — Co-designs a DSL with a multi-agent system for 3D program synthesis; relational operators let agents compose geometry instead of emitting raw code.

Research is converging on guardrails for agent tool use: memory bounds, reasoning gates, and structured topologies are the new reliability primitives.

⚙️ Ship — 2 items

📦 Claude Files API and Admin API user management GA — Claude Files API and Enterprise Admin user-management endpoints go GA — beta headers no longer required; file expiration and pagination are now standard.

📦 Anthropic SDK 0.120.0 — Adds managed agents web search config and self-hosted sandbox memory — new knobs for agent search and sandbox persistence.

Platform maturity day: Claude's Files and Admin APIs drop beta headers while the SDK adds managed-agent search and sandbox memory. Operators can start removing beta flags.

💬 Community — 2 items

💬 Claude Code automates neuroscientific data analysis — Eli Sennesh: Claude Code automates most of his one-time job as a neuroscientific data analyst — "which is good, because I super hated that this was an entire job." 61 likes.

💬 Show HN: HarnessRouter: Unified interface for agent harnesses — Canonical API for running Codex, Claude Code, Hermes, and other managed agent harnesses as your product backend. 10 points, 14 comments.

Harness abstraction is the theme: a unified API across Codex/Claude Code/Hermes, and Claude Code already replacing a whole analyst job.

📚 Read — 4 items

📚 Co-RL: Unsupervised Reasoning Emerges from Diverse Cohort in Multi-agent RL — Cooperative multi-agent RL where peer-derived rewards replace human labels — reasoning emerges without supervised data. Relevant if you're exploring multi-agent training without reward engineering.

📚 SPADE: Self-Play in Adaptive Synthetic Executable Environments — LLM writes its own executable training environments and learns via self-play — environment scaling without hand-built simulators. Points at self-improving agent training loops.

📚 FM-Bench: A Benchmark for Long-Horizon Management with Competing Agents — Benchmarks LLM agents managing a football club over 20 simulated years with competing agents — long-horizon decision-making under competition. Useful eval for agents that need to sustain strategy, not just answer.

📚 Emergence of Agentic AI: A Review on Evolution, Background, Working Principles, Applications, Adoption Factors, and Future Research Directions — Broad survey of agentic AI — evolution, working principles, adoption factors, future research directions. Useful orientation if you're evaluating where agent frameworks are heading.

Multi-agent RL is moving from hand-crafted rewards and simulators to self-generated ones. If you're building agent training pipelines, the bottleneck is shifting from data to environment design.

💬 Community — 2 items

💬 Bounded Agents: Delegation Security for Multi-Agent AI Systems — Agentic Principal Chain enforces session-aware authorization to block harmful action combinations and delegation abuse in multi-agent systems.

💬 Looped Language Models Improve Compositional Tool Calling — Looped LMs use recurrent computation for multi-step tool use, with adaptive inference balancing accuracy and compute.

Security and tool-calling reliability are the two threads the community is pulling on this week — both papers point at concrete failure modes operators will hit.

⚙️ Ship — 4 items

📦 OpenClaw release-publish/bfa7c3d38e6d — OpenClaw release-publish/bfa7c3d38e6d — no release notes surfaced in the feed; check the tag for operator-facing changes.

📦 Codex 0.149.0 — Codex 0.149.0 adds an interactive `codex agents` dashboard for managing tasks, `codex queue` to message existing sessions, and `codex doctor` diagnostics for endpoint/network/desktop issues.

📦 Antigravity Google updated — Antigravity changelog updated — feed excerpt unreadable; check the changelog for operator-facing changes.

📦 Claude platform release notes — August 19, 2026 — Claude API GA's computer use (`computer_toolset_20260801`) with batch actions and zoom, and launches a browser use tool (`browser_toolset_20260801`) that reads the page's accessibility tree instead of screenshots.

Browser automation is consolidating: Claude's new browser tool reads page structure directly, and Codex adds a task dashboard and queue. Operators should re-evaluate screenshot-based browser agents.

📚 Read — 5 items

📚 Position: Multi-Agent Systems Should Prioritize Concurrency Control — Position paper argues multi-agent systems should treat concurrency control as a first-class primitive — worth reading before your next orchestration layer.

📚 EnvHarness: Awakening Static Worlds for Agent Learning — EnvHarness + EnvRigger reshape static environments via programmable plugins to target agent weaknesses — a practical route to harder RL training worlds.

📚 PolicyGuide: From Guarding One Action to Guiding the Whole Workflow for Policy-Compliant LLM Agents — PolicyGuide shifts from guarding individual actions to guiding entire workflows for policy-compliant LLM agents — useful if you're enforcing compliance across multi-step runs.

📚 Self-Evolving Agents as Dynamic Graph Transformation: A Survey and New Perspective — Survey reframes self-evolving agents as dynamic graph transformations — a useful mental model for tracking how agent capabilities drift over time.

📚 FinSkillBench: Evaluating AI Agents and Domain Skills for Investment Management — FinSkillBench evaluates AI agents on domain skills for investment management — a benchmark for anyone building finance-facing agent workflows.

Research day: the throughline is making agents more controllable — policy compliance, concurrency, and harder training environments.

⚙️ Ship — 3 items

📦 Claude platform release notes — August 19, 2026 — Claude API: computer use tool goes GA (`computer_toolset_20260801`) with batch actions and per-member config; new browser use tool (`browser_toolset_20260801`) for driving a browser in a viewport. Both change how agents interact with UIs.

📦 DeepSeek V4-Flash-Vision-Exp: Multimodal API Live — DeepSeek ships V4-Flash-Vision-Exp — multimodal API now live, adding vision to the Flash line. Relevant for agents that need image understanding without the full-size model cost.

📦 Antigravity Google — Faster project switching, redesigned Customizations settings, and bug fixes — Antigravity changelog: faster project switching and redesigned Customizations settings. Minor but operator-facing for anyone using Antigravity daily.

Tool-use surface is expanding fast: Claude adds browser and computer toolsets, DeepSeek adds vision. Agents that can see and click are becoming table stakes.

💬 Community — 3 items

💬 SWE-bench Science: Can Coding Agents Resolve Engineering Tasks in Science? — New benchmark extends SWE-bench to scientific codebases; agents struggle with domain-specific failures and scientific guidance helps only sometimes.

💬 MemTrapBench: Benchmarking Cognitive Traps in LLM Memory Use — Retrieved memories can induce reasoning errors and belief distortions in LLMs; an inference-time strategy helps avoid these traps.

💬 SkillGate: Training In-Policy Skill Selection in Long-Horizon Agents — SkillGate fixes selector credit starvation in agent skill selection by separating outcome credit for execution tokens from local advantage.

Three papers converge on the same theme: agent reliability is still fragile — memory can mislead, skill selection needs better credit assignment, and coding agents stumble on domain-specific tasks.

📚 Read — 3 items

📚 FreeToken — A paper proposing an edge-native MoE serving system to run models like DeepSeek-V4-Flash on consumer hardware.

📚 FIXING Opus 5 — A video critique arguing that prompt engineering remains essential for model optimization.

📚 StreamOPD — A post-training framework using cue-gating and distillation to improve streaming video understanding.

💬 Community — 1 item

💬 Show HN: OzBrain, a shared brain for knowledge between agents and your team — Show HN: OzBrain pitches a shared knowledge layer between agents and humans — agent-first chat as the primary interface, dashboards out. Early, but the thesis is the debate.

The agent-first interface thesis keeps surfacing in community builds; whether shared knowledge layers become the default is the open question.

⚙️ Ship — 4 items

📦 Claude Python SDK v1.0 — August 20, 2026 — Python SDK v1.0 — breaking: HTTP layer moves httpx→httpx2, temperature/top_p/top_k dropped from Messages, legacy Text Completions removed, async `.with_raw_response` now needs `await response.parse()`. Migration guide has before/after snippets for every change.

📦 Antigravity — Remote Control, faster project switching, and more — Remote Control ships — drive Antigravity from another device, plus faster project switching. Google's agentic IDE keeps landing operator-facing moves.

📦 OpenClaw release (tag 1a3aa6a36af7) — New OpenClaw release tagged. Changelog not surfaced in this cycle — check the release page for runtime changes before upgrading.

📦 xAI API release notes — August — xAI release notes updated for August — Grok 4.6 is the current model line. Specific changes not surfaced in this cycle; check the page for model or pricing moves.

Breaking-change day for the Claude stack. If you're on temperature/top_p/top_k params or the legacy completions API, the v1 migration guide is your afternoon — before the next release compounds it.

📚 Read — 5 items

📚 Repo0: Design-Driven Zero-to-All Code Generation — Repo0 generates complete repositories from natural language by evolving a dual-graph architecture as it codes — a step past single-file coding agents.

📚 FACET: Preserving Source Intent and Executable State in Terminal Task Synthesis — FACET builds executable terminal tasks by grounding instructions, solutions, and verifiers in a shared representation that preserves source intent — scalable supervision for training terminal agents.

📚 Hallucination as a Feature, not a Defect: Evaluating a multi-agent architecture to transform speculative language-model outputs into testable scientific hypotheses — Reframes hallucination as a feature: a multi-agent architecture turns speculative LM outputs into testable scientific hypotheses, with evaluation of the pipeline.

📚 Position: Collusion Risks Among AI Reasoning Agents Justify Certification Requirements for Making Market Decisions — Position paper argues collusion risks among AI reasoning agents justify certification requirements before they make market decisions — a governance angle on multi-agent deployment.

📚 Test Agent Changes with LangSmith Preview Builds — LangSmith preview builds let you test agent changes against production traffic before rollout — staging for agent pipelines, not just code.

Coding-agent research is moving from single-file completion to whole-repo architecture generation, while terminal-agent training gets executable supervision. The next harnesses will be judged on repo-level coherence, not line-level accuracy.