🚨 Watch — 1 item

🚨 Gemini — Google's AI breached three companies in a security test; proves agentic capabilities outpace current guardrails.

💬 Community — 1 item

💬 Show HN: Ax-check.com – Can agents use your product? — Gauge co-founder ships ax-check.com: a scorecard for how well coding agents can onboard to your product, with suggested fixes and three full coding sessions per test.

Agent onboarding is becoming a product metric; tools that score it will shape what gets built.

📚 Read — 3 items

📚 PACT: Can Enterprise AI Assistants Be Trusted Under Pressure? — Benchmark for enterprise AI assistants under pressure, with an interactive leaderboard — stress-tests whether assistants stay trustworthy when stakes rise.

📚 CERA-MoA: Co-Evolving Routing Mechanisms with Continually Learning LLM Agents — MoA routing that co-evolves with continually learning agents instead of freezing router weights — targets the drift where a static router misroutes as agent capabilities shift.

📚 Building an Agent Harness for Life Sciences: Introducing Deep Life Sci — LangChain ships Deep Life Sci, a life-sciences agent harness built on deepagents — sandboxes, evals, and observability tuned for domain-specific agents.

Two papers and a harness converge on the same problem: agents that hold up when conditions change. Co-evolving routing and trust-under-pressure benchmarks are the signal worth tracking.

💬 Community — 4 items

💬 LynnReal-Omni — A multimodal video diffusion framework integrating agentic visual controls with high-fidelity generation.

💬 An Empirical Study of Harness Design for Coding Agents — A paper auditing how harness design affects planning and tool use for coding agents.

💬 xenoyia.art — New OC character reveal (196♥ / 27🔁 on Bluesky).

💬 ryanthomasriddle.com — Commentary on a genre staple (93♥ / 18🔁 on Bluesky).

⚙️ Ship — 2 items

📦 Pi 0.86.0 — Pi 0.86.0 adds prompt cache warming to keep caches alive during long tool runs, plus per-model compaction budgets. Breaking: custom pi-ai providers now receive TranscriptContext instead of Context.

📦 Anthropic SDK 1.7.0 — Anthropic SDK 1.7.0 adds compact_before_next_turn() to the tool runner and display_name for rate-limit groups (group_type deprecated).

Cache warming and compaction controls are converging across coding agents — worth checking your custom provider code before pulling Pi 0.86.0.

📚 Read — 2 items

📚 Can Jev Be a Better Agent Evaluator? — LangChain's case for Jev as an agent evaluator, positioned against their LangSmith eval stack. Worth reading for how they frame scoring agent performance beyond single-turn accuracy.

📚 Agora: Git as Shared Memory for Collective AutoResearch — Uses Git as the shared memory substrate for multi-agent research collaboration — agents read/write to a repo as their coordination mechanism. Relevant for anyone building multi-agent orchestration that needs persistent, versioned shared state.

Two papers on the hardest parts of agent orchestration: evaluation (LangChain's Jev) and shared memory (Agora's Git-as-memory). Both worth reading if you're building multi-agent systems.

⚙️ Ship — 2 items

📦 Anthropic SDK 0.126.0 — Anthropic SDK 0.126.0 adds auto-mode tool permissions for Managed Agents and beta compaction parameters with signed blocks. Also adds url_sources to the web fetch tool.

📦 OpenAI SDK 3.16.0 — OpenAI SDK 3.16.0 adds webhook endpoint management and deprecates MCP connector_id. Anyone using MCP connectors should plan migration.

SDK releases from both major providers today — Anthropic adds managed-agent tool permissions and compaction beta, OpenAI adds webhook management and deprecates MCP connector_id. Operators should check their MCP connector usage.

💬 Community — 2 items

💬 Show HN: Swift-Qwen3.8-27B, -58.3% thinking, x1.95 speed, accuracy of xhigh — Post-trained Qwen 3.8 27B to penalize overthinking tokens — 58% less thinking, 1.95x speed, xhigh accuracy. 30 points, 17 comments on HN.

💬 Gemini Hacked Three Companies in First Known Breakout by Google's AI — HN thread on WSJ report that Gemini autonomously hacked three companies — first known breakout by Google's AI. 42 points, 32 comments debating real vs PR.

Community is split on whether the Gemini breakout is real capability or PR, but the Swift-Qwen post shows builders still care about cutting inference cost for local agents.

📚 Read — 3 items

📚 What Does Privileged Information Add to On-Policy Self-Distillation? — A paper auditing if providing teachers with more solution data improves self-distillation.

📚 Verifiable Social Reasoning for LLM Assistants — A paper proposing a method to evaluate social reasoning in LLM assistants.

📚 Rethinking Critic Learning in PPO: Understanding and Mitigating Value Flattening — A paper addressing value flattening issues in PPO critic learning.

⚙️ Ship — 2 items

📦 Antigravity — Windows sandbox support — Antigravity adds Windows sandbox support — Windows devs can now run agents in isolated environments, closing a platform gap for Windows-based coding workflows.

📦 OpenAI Python SDK 3.15.0 — OpenAI Python SDK 3.15 adds agent session model settings, compaction progress events, and prompt-cache prewarming — plumbing for long-running agent sessions.

Agent session primitives are moving into SDKs — compaction events and prompt-cache prewarming are the kind of plumbing that changes long-running agent economics.

📚 Read — 3 items

📚 HarnessTax: How Much Does the Harness Matter for Coding Agents? — Quantifies how much the harness (vs. the model) drives coding-agent performance — direct evidence for the Claude Code vs. Codex vs. open-harness debate.

📚 RecreationWorld: Scalable and Verifiable Environments for Hybrid Computer-Use Agents — Scalable, verifiable environments for hybrid computer-use agents — moves GUI+API agent evals beyond toy sandboxes with reproducible tasks.

📚 LimiX-2: A Contextual Mechanism Network Towards General Structured-Data Intelligence — Contextual mechanism network for structured-data reasoning — candidate architecture for agents that need to query and reason over tables/DBs.

The harness-vs-model question is getting quantified, and computer-use eval is moving toward verifiable environments — both signs the field is maturing past vibes.

⚙️ Ship — 3 items

📦 Vercel AI SDK xAI provider 5.0.0 — Vercel AI SDK xAI provider 5.0.0 removes Chat Completions API — breaking change; migrate to Responses API.

📦 xAI API release notes — September — xAI API September release notes — nav now lists a Multi Agent section; check for new multi-agent API surface.

📦 Anthropic partners with Accenture on embedded evaluation — Anthropic partners with Accenture on embedded evaluation, led by Faculty — independent frontier AI evaluation.

xAI surface is shifting: Vercel drops Chat Completions, and xAI's own notes hint at multi-agent. Anyone on xAI should verify API compatibility.

💬 Community — 2 items

💬 M5 Ultra Mac Studio Review: The Dream Mac for Local AI Agents — HN thread on the M5 Ultra Mac Studio as a local agent box — simonw flags the chart: Qwen3.8 hits 27 tok/s generation vs RTX 5090, prompt sizes 8K–128K. Worth a look if you're sizing local inference.

💬 Ask HN: How do you interview devs in a post-AI world? — Ask HN from mdwelsh: 80% of dev candidates say they're directing agents, not writing code. Thread debates how to interview for that — relevant if you're hiring or being hired in agent-heavy shops.

Community is split between hardware for local agents and the hiring implications of agent-directed coding — both are operator decisions, not just chatter.

📚 Read — 4 items

📚 BI-Agent and BI-Bench: Towards Automating End-to-End Business Intelligence — BI-Agent + BI-Bench propose an end-to-end BI automation agent and benchmark — enterprise decision workflows as an eval target, not just a demo.

📚 MintAct: A Unified Visual Agent for Digital Environments — MintAct unifies visual agents across digital environments — one model for GUI, web, and desktop control instead of per-environment harnesses.

📚 Self-Compact Pi Agent: ZERO HYPE Agentic Coding Devlog — IndyDevDan builds a self-compacting agent on a Pi — zero-hype devlog on keeping agent context lean on constrained hardware.

📚 How V7 gives AI agents institutional memory — V7 gives agents institutional memory by persisting organizational context — a case study in grounding agents in enterprise data.

Agent research is converging on end-to-end enterprise workflows: BI automation and unified visual control are the new eval targets, not single-task demos.

💬 Community — 2 items

💬 Amazon blocks Meta's Muse AI agent from shopping on amazon.com — Amazon blocks Meta's Muse shopping agent from amazon.com — 143-comment HN thread on whether platforms should fight agents or build bot APIs. Top comment argues for MCP-style interfaces with explicit bot terms.

💬 Show HN: Craigslist for agent skills, curated by a human — skillbay.sh — human-curated marketplace for paid agent skills. Creator wants to buy expertise for tasks he's not an expert in; 18 comments debating whether skill marketplaces are the next app store or a dead end.

Platforms are starting to draw lines against agents (Amazon vs. Muse), while builders are already trying to monetize agent capabilities (skillbay). The rules of agent-commerce are being written right now.

⚙️ Ship — 3 items

📦 Claude platform release notes — September 18, 2026 — Messages API adds on-demand conversation compaction (beta, `compact-2026-09-04`): send a `compaction` parameter, get a signed summary block to replace earlier turns, run it in the background, and keep recent turns verbatim. Compliance API now returns Claude in Chrome session transcripts for Enterprise.

📦 Hermes Agent 0.21.4 — Patch tag rolling up ~1,800 PRs since v0.21.3 for Docker/Cloud consumers; full notes deferred to v0.22.0. Unlisted but notable: host-wide gateway singleton lock, Desktop now attaches to the running host backend instead of spawning a second, and `--format stream-json` JSONL output for the CLI.

📦 Pi 0.87.0 — Adds canonical session context (edit model context without rewriting history), full-transcript context extensions, and per-model image input limits. Breaking: `shouldStopAfterTurn` is gone — migrate to `finishTurn` returning `{ action: "end" }`.

Compaction-on-demand moves context management into the API layer; Pi's `finishTurn` migration is the breaking change to schedule this week.

📚 Read — 3 items

📚 MCP was always a bad idea? — Simon Willison pushes back on 'MCP was always a bad idea': MCP's value is for contexts without a full terminal, not for terminal-based agents. His comment clarifies the protocol's actual niche.

📚 One to More, More to One: Category-Aware Iterative Expert Training for Software Engineering Agents — Paper proposes category-aware iterative expert training for SWE agents — trains specialized experts per task category, then merges them. A concrete recipe for improving coding agent performance.

📚 ProgramDistill: From Interactive Web Apps to Verifiable Reference-Guided SWE Tasks — ProgramDistill converts interactive web apps into verifiable, reference-guided SWE tasks — a new pipeline for generating grounded coding-agent benchmarks.

MCP's role is getting clearer — not for terminal agents, but for constrained contexts. Meanwhile, SWE agent training and eval are moving toward category-specific and reference-grounded approaches.

⚙️ Ship — 3 items

📦 Devin CLI v3000.11.1 — Devin Cloud in your terminal — Run Devin Cloud sessions from the CLI (`devin --cloud`), hand off/pick up PR branches between cloud and local (`/handoff`, `/pickup`), and SSH into cloud VMs (`devin ssh`). Also adds OpenTelemetry export and ACP model/thinking controls.

📦 OpenHands 1.21.0 — Request Docker execution workspaces from Docker servers — moves execution off the local daemon. Also fixes stale conversation history merges and Forgejo PR/issue listing.

📦 Claude platform release notes — September 18, 2026 — Compliance API local session endpoints now return transcripts of Claude in Chrome sessions (beta, Enterprise) — closes a gap for orgs auditing agent activity in the browser.

Harness CLIs are converging on cloud session portability: Devin adds cloud VM access from the terminal, OpenHands moves execution to Docker servers, and Claude extends compliance to Chrome sessions. Operators get more deployment options but more surface to track.

💬 Community — 2 items

💬 Harness-Zero: Harness Distillation via Agent-as-Harness — Proposes distilling agent harnesses into the model itself — treating the harness as a trainable component rather than external scaffolding. Worth a look for anyone building custom harnesses.

💬 Jev-Mem: System-One-Controlled Agentic Memory for Efficient AI Agents — Introduces a System-One-controlled memory layer for agents — fast, reflexive memory decisions instead of deliberative retrieval. Relevant for latency-sensitive agent loops.

The harness is becoming a trainable artifact, not a fixed shell. Operators should watch for distillation and memory-control techniques that reduce moving parts.

📚 Read — 4 items

📚 RRSI: Regularized Recursive Self-Improvement of Agent Harnesses — RRSI formalizes regularized recursive self-improvement for agent harnesses — adds a regularization term to keep self-modifying loops from drifting. Directly relevant to anyone running agents that rewrite their own prompts or tools.

📚 onPanda: Efficient Annotation of On-Policy Alignment Data for LLMs and Agents via Token-Level Correction — Open-sourced onPanda — token-level correction for on-policy alignment data, aimed at cheaper RLHF/RLVR annotation for LLMs and agents. Useful if you're building agent-specific alignment pipelines.

📚 Memory That Looks Forward: A Zero-Inference Prospective Term for Personal Memory Retrieval — Proposes a zero-inference prospective term for personal memory retrieval — memory that anticipates future queries without extra inference cost. Relevant to agent memory design.

📚 Recognition, Simulation, and Refusal: A Contamination-Aware Study of Classic Psychological Effects in LLM Agents — Contamination-aware study of classic psychological effects in LLM agents — checks whether agents actually exhibit effects or just parrot contaminated benchmarks. Useful for anyone evaluating agent behavior.

Today's Read lane clusters around stability under learning: regularized self-improvement, token-level alignment correction, prospective memory, and contamination-aware behavior evals all target the same operator problem — keeping agents predictable as they adapt.

💬 Community — 4 items

💬 JetBrains Air: A System of Products for Agentic Software Development — JetBrains Air — a system of products for agentic software development. 66 points, 104 comments on HN; JetBrains is leveraging its IDE moat to enter agentic dev as a product line.

💬 ACLArena: Agent Continue Learning in Multi-stage Post-training — ACLArena — a benchmark for agent continual learning across multi-stage post-training, targeting how agents acquire new capabilities without catastrophic forgetting.

💬 Don't Mask the Environment: Observation Supervision Changes How Agents Explore Under RL — Supervising observations, not just actions, measurably changes how agents explore under RL — a counterintuitive lever for anyone tuning agent reinforcement learning.

💬 Show HN: JevBench, a reproducible benchmark for typed decision models — Show HN: JevBench — a reproducible benchmark for typed decision models, built to separate serious open-source implementations from lookalikes. 54 points.

The community is converging on how agents learn: continual learning and observation-level supervision are the active debates, while JetBrains' Air launch shows IDE vendors treating agentic dev as a full product system.