Pulse (last 3d) · Research 111 · Agents 71 · Products 55 · Industry 42 · Models 37 · Legal 35 · Open Source 25 · Policy 18 · Infra 15 · Enterprise 14

Trending · OpenAI 24 · Claude Code 13 · Codex 10 · FDA 9 · Anthropic 8 · Claude 6 · Hugging Face 6 · Nvidia 6 · Meta 5 · The Verge 5 · MCP 4 · Microsoft 4

Top story

Today's agent research reaches one conclusion: the failures that matter now come from what an agent trusts. That can be a stale memory record, a persisted rule or a scorer used outside its applicability domain, and the fix is versioned provenance with invalidation, not a better model. Tooling vendors are reaching the same place from the other side by gating credentials, actions and spend outside the model, while open weights trail the closed frontier by about a month and 2-bit quantization makes them cheap to host. Model choice is becoming the commodity decision for clinical agents, and the record of what each decision depended on is the part that stays scarce. arxiv.org | latent.space

The Shortlist

  • Tying agent memory to dependency versions cuts unsafe publication from 62% to 0.3%. arxiv.org
  • Throwing away agent guidance after each run beats keeping it, because stored rules regenerate their own errors. arxiv.org
  • MAGI ran on 9 real pharma campaigns, and the scorer's applicability domain, not the agent, limited what it achieved. arxiv.org
  • Kimi 3.1 Pro is Fable-5.1-class with open weights, putting open models about one month behind closed ones. latent.space
  • A 2-bit DeepSeek-V4.1-Flash reportedly matches the full-precision model, and llama.cpp support has landed. i.redd.it
  • 1Password now handles agent logins with approval on the user's device, and the agent never sees the password. producthunt.com
  • OpenAI is reportedly acquiring Orchestra Bio. x.com

Also Noted

  • OpenAI acquires Orchestra Bio - reported in a roundup that also covers AI-designed phages used in a patient (single source). x.com
  • Claude Fable 5.5 leak - sightings on Azure, GitHub and Cursor point to Anthropic's biggest 2026 launch (single source). x.com
  • Meta byte-level models - they start worse than token-based models, but scaling fits predict they overtake them. i.redd.it
  • Raschka reasoning course - a full video course on building reasoning models from scratch. magazine.sebastianraschka.com
  • Cerebras trillion-parameter serving - how Cerebras runs trillion-parameter frontier models on its wafer-scale systems. cerebras.ai
  • Strata/Basalt - runs Qwen3.8 Flash Next IQ3_S at 250 tok/s on 16GB VRAM plus 55GB RAM. reddit.com
  • 8GB GPU, 262K context - a 3.5 BPW Qwen coder runs at 70 tok/s through llama.cpp. reddit.com
  • RTX 30/40 unlock tool - reportedly doubles prefill speed, and used-card prices are rising. wccftech.com
  • RTX 5090s in Chinese labs - their dies and 32GB VRAM are being pulled off for AI servers. wccftech.com
  • memt-1.0 - agent memory built on gated DeltaNet-2 and mixture-of-storage, claiming a 66x smaller footprint. i.redd.it
  • OAK.memory - agent memory across sessions using vector search plus graph traversal, through MCP or locally. producthunt.com
  • GPT-X3-150M - a GDN2/GQA hybrid that takes first place on the Open SLM Leaderboard from SmolLM2-135M. huggingface.co
  • Reactive Optimizers - the optimizer's source code is rewritten while training runs. xingjianbai.com
  • EPFL neural chip - a chip that decodes interactions between brain regions, tested on iEEG data from psychiatric disorders. actu.epfl.ch
  • UK AI Biosecurity Lab - the UK commits £100m to a lab it calls a "world first". bbc.co.uk
  • Seclai - traces every agent run, applies governance rules written in plain English and runs the same evals locally, in CI and in production. producthunt.com
  • LLM Test Server - a local, repeatable stand-in for 13 LLM provider APIs that shows where every token goes. producthunt.com
  • Truvec - tests embeddings, chunking, hybrid search and rerankers on your own documents before you change production. producthunt.com
  • Looper - an open-source agent given one metric to improve, which asks the team before anything goes out. producthunt.com
  • Supermux Core - one TypeScript API for sessions, events and permission prompts across six coding-agent CLIs. producthunt.com
  • Inyon - answer coding agents running in tmux from your phone over an encrypted tunnel; MIT-licensed. producthunt.com
  • DotMD - review Markdown written by agents like a pull request, and the agent fixes each comment over MCP. producthunt.com
  • Prompt Deck - puts memory files, slash commands and sessions from 11 coding agents in one window. producthunt.com
  • TaskSquad - a shared workspace where people work with Claude Code, Codex or Gemini agents. producthunt.com
  • Pitlane - builds and debugs Xcode projects without opening Xcode, and a local API lets agents fix and rebuild. producthunt.com
  • sideOS.ai - a voice-driven agent that operates your Mac and can run fully offline on local models. producthunt.com
  • SerieX - open-source bookkeeping and payroll that works offline, with a connector for an AI assistant. producthunt.com
  • Grok vs OpenAI party test - when forced to choose, Grok picks Republicans and OpenAI's model picks Democrats. techcrunch.com
  • Creative-sector job losses - looks at how much of the roughly 200k lost US media, film and arts jobs is due to AI. apricitas.io
  • Pennsylvania AI ad labels - the state will require labels on AI-generated ads. theverge.com
  • Block voice model - Block's new speech model sounds close to human. techcrunch.com
  • Live music transformer jam - the Anticipatory Music Transformer jams live, steered by logit biases from key, chord and tempo tracking. reddit.com
  • Browser rhythm game - four small neural nets analyze songs inside the browser tab; open source. v.redd.it

Agent errors now start in what the agent trusts, not in how it reasons

In each item the failure sits in stored state or an outside scorer, and each fix is provenance plus invalidation.

Stored agent guidance brings back its own mistakes, even after it is deleted Wang, Panaganti and Zhou show that a faulty stored rule shapes an agent's runs, and new rules built from those runs repeat the error after the original is deleted (preprint, unreplicated). Throwing guidance away after each run did better than keeping it, which challenges the common design of self-improving agents that build up lessons across clinical workflows. arxiv.org

Tying memory verdicts to dependency versions cuts unsafe publication from 62% to 0.3% Tool-using agents published 62% of objects whose failure traced back to a stale memory record. Pinning each verdict to exact dependency versions when it is written, and checking for invalidation when it is read, cut that to 0.3% with 37% fewer verification calls (preprint, unreplicated). It applies to agent beliefs the same lineage discipline regulated data pipelines already need. arxiv.org

MemTX supplies the transactional spec that lineage-based memory was missing MemTX groups memory writes into isolated transactions that record evidence, permissions, provenance and validity for each entry. It holds back irreversible tool calls while beliefs are still uncommitted, and retracting one belief also retracts the ones that depend on it (preprint, unreplicated). Together with dependency tethering, the stale-memory gap now has two designs teams can build. arxiv.org

MAGI's lead-optimization ceiling is set by the scorer, not the agent Across 9 real pharma lead-optimization campaigns, MAGI's results were capped by the range the scoring model is valid for, not by what the agent could do (preprint, unreplicated). For drug-discovery agents, better property predictors with wider coverage now pay off more than a stronger planner. arxiv.org

Agent control is becoming its own product layer at the point of action

Each launch takes authority away from the model and gives it to a gate that approves credentials, actions or spending before anything runs.

1Password handles agent logins without exposing credentials When an agent hits a login, 1Password for Agents fills the password or one-time code after the user approves on their own device, and connecting an agent does not open the whole vault. It is included in Individual and Family plans, with Meta Muse and Grok Bot rolling out, so per-action human approval becomes a consumer default instead of something enterprises build themselves. producthunt.com

Action gates for agents now ship as plain-English rule engines Fluxix checks each agent action against rules written in plain English, blocks risky operations, asks for human approval and logs every decision. Plango uses an on-device security model to control coding agents' access to files, secrets and the network. Both treat the agent as untrusted and keep the audit log outside it, which is the stance a HIPAA review expects. producthunt.com | producthunt.com

Spending caps now block tool calls from outside the model Peras refuses a coding agent's next tool call once weekly usage passes a set limit. agent-usage-guard reads live 5-hour and weekly limits, asks before subagent fan-out, heavy-context prompts and retry loops, and refuses when nobody answers. Both are MIT-licensed and work through harness hooks, so the model cannot override the cap. producthunt.com | producthunt.com

Open weights trail the closed frontier by about a month, and leaderboards no longer settle the ranking

The gap is now small enough that real-world testing and release policy separate the models more than benchmarks do.

Kimi 3.1 Pro brings Fable-5.1-class capability to open weights, without cyberattack training data Moonshot released Kimi 3.1 Pro as open weights, with 104B active out of 2.8T total parameters, claimed to be Fable-5.1-class, plus a Highspeed API that runs 2 to 3 times faster (vendor benchmark). Reuters reports that cyberattack training data was left out of the release. One user found the model, ranked #4 on Artificial Analysis, falls behind DeepSeek 4.1 Flash and MiMo 2.6 in real use (single source). latent.space | reuters.com | reddit.com

Meta's MSL ships a Fable-class frontier model, and keeps it closed Meta Superintelligence Labs released Muse Wonder and Wonder Full as closed-source frontier models at the Fable tier. With Meta moving from open Llama weights to closed models, Chinese labs become the main source of frontier-class open weights for anyone who must self-host models on patient data. latent.space

Five Qwen3.8-27B models working together roughly match Fable 5 at coding The GVS5H harness runs five Qwen3.8-27B models together and comes close to Fable 5 on an older LiveCodeBench-style benchmark (single source). On narrow tasks, coordinating mid-size open models is starting to replace frontier scale, though the old benchmark limits how far the result carries. github.com

Compression is moving near-frontier models onto hardware already on the desk

Quantization, runtime work and vendor edge releases each cut the memory that useful local inference needs.

ISTA's 2-bit DeepSeek-V4.1-Flash matches the full model, and llama.cpp now runs it ISTA's GSQ-RCO 2-bit quantization of DeepSeek-V4.1-Flash reportedly scores the same as the unquantized model (single source), and a llama.cpp pull request from am17an adds support for the model. If the result replicates, hosting a strong MoE on-prem needs a fraction of the memory it did before. i.redd.it | github.com

Apple silicon becomes a credible target for offline inference ANEMLL Forge 0.2.1 runs Qwen3.8-27B on the M6 Neural Engine with up to 192K context, an INT8 KV cache and a smaller draft model. macOS 27 also reportedly ships an offline fm terminal LLM with an OpenAI-compatible local endpoint (single source). Fully offline inference on a 32GB Mac Mini makes data-residency arguments for internal tools simpler. reddit.com | x.com

NVIDIA's Nemotron-3.5-Bolt-E4B fits vision into 6GB NVIDIA released Nemotron-3.5-Bolt-E4B, a vision-capable edge model trained on Gemma 4 that runs in 6GB of RAM, and Unsloth has published GGUFs. Multimodal inference at this size can run on lab instruments and devices, so images never leave the bench. x.com

Dropped

15 items: no content, off-topic hobby projects, promotional posts, help requests or opinion pieces with no claim.