Every post I’ve considered, picked or not, with one line on each, starting with Issue #3. Narrow it to a month, quarter or year, or by source and interest. “Not reviewed in full” means it was judged from a skim.
57 of 57 posts
Anthropic's method for building safety evals people can trust, from choosing what to measure to reporting uncertainty.
Summary in Issue #4Building agentsAdvisingMarkets & safe use
Cloudflare open-sources Kitesurf, its lightweight browser for agents that runs on Workers. Not reviewed in full.
Building agents
Holistic AI adds runtime controls for what agents can access, run and spend. Not reviewed in full.
Markets & safe use
llama.cpp reaches 1.0, with a commitment to versioned, stable APIs. Not reviewed in full.
Inference stack
LangChain announces a $300M Series C at a $3B valuation. Not reviewed in full.
Markets & safe use
Google's fast, low-cost Gemini 4 model, pitched as the worker under its frontier models.
Summary in Issue #4Choosing modelsAI ecosystem
Google tests its agent anomaly detector on five outside agent-security benchmarks.
Building agentsMarkets & safe use
The controls OpenAI uses to run hundreds of always-on agents internally.
Summary in Issue #4Building agentsMarkets & safe use
A conference report on what engineers are building with System One models such as Jev.
Choosing models
Google releases A2UI 1.0, an open protocol for agents to answer with native interface components. Not reviewed in full.
Building agents
Argues that when AI writes the code, tests should become customer-readable data files the AI can't quietly rewrite.
Building agents
vLLM's work on serving multimodal models with the encoder split from prefill and decode.
Inference stack
A Duolingo talk on agents writing and merging 225,000 lines of tests. Not reviewed in full.
Building agents
Google makes Gemini 4 Carbon, its strongest model, generally available. Not reviewed in full.
Choosing modelsAI ecosystem
DeepMind's agents formally verify key security properties of the Signal protocol's code. Not reviewed in full.
AI ecosystemBuilding agents
Four companies share how they build voice agents on GPT-Live. Not reviewed in full.
Building agents
Hugging Face's common interface for decision models, which return probabilities instead of generating text.
Summary in Issue #4Choosing modelsInference stack
How LangChain's own sales agent produces consistent sales documents. Not reviewed in full.
Building agentsAI business
How Workday built an internal agent platform on LangGraph and LangSmith.
Building agentsAdvising
A hands-on review of NaiveAI's very fast open coding model. Not reviewed in full.
Choosing models
Resend's CEO on getting picked when coding agents choose the vendor.
Summary in Issue #4AI businessBuilding agents
How durable execution lets agents run for weeks, and which failures the harness still has to handle.
Summary in Issue #4Building agents
OpenAI research on how AI is affecting freelance work. Not reviewed in full.
Markets & safe useAdvising
A Microsoft talk on treating sub-agents as a distributed system. Not reviewed in full.
Building agents
Claude designs three working electronic devices, from circuit boards to enclosures and firmware. Not reviewed in full.
AI ecosystem
Three Spotify product managers on how the AI PM role differs by what you build. Not reviewed in full.
Advising
An explainer on improving agents by writing evals first. Not reviewed in full.
Building agents
A deep dive into what frontier labs buy from Handshake's AI training-data business. Not reviewed in full.
Markets & safe useAI business
OpenAI's system for tracking and stopping model workloads across its research infrastructure. Not reviewed in full.
Markets & safe use
Benchmarks for whether coding agents can recover from failures and undo what they did.
Summary in Issue #3Building agentsMarkets & safe use
Explainer on why the same model runs at different speeds on different chips, centred on the KV cache.
Summary in Issue #3Inference stack
Release notes for LangChain's Managed Deep Agents: browser use, background subagents and Slack thread handling.
Building agents
A Modal talk on building infrastructure for agents. Not reviewed in full.
Building agentsInference stack
A Fieldguide talk on making financial agents show how they reached their answers. Not reviewed in full.
Building agentsAdvising
Cohere's experiments on teams of coding agents: when a lead plus workers beats a single agent, and at what cost.
Summary in Issue #3Building agentsChoosing models
How Anthropic designs Claude's memory: session notes plus an offline pass that corrects them against past transcripts.
Summary in Issue #3Building agents
Guest post from Tübingen researchers on how AI is starting to help build scientific models, starting from neuroscience. Not reviewed in full.
AI ecosystem
Google measures how often coding agents detour into stale TODO comments, and what those detours cost.
Summary in Issue #3Building agentsAdvising
Ai2's ToyBlocks, a simple simulated world for studying what robot models learn and why they fail. Not reviewed in full.
AI ecosystem
The Open ASR Leaderboard adds private test data for Egyptian and Nigerian English accents. Not reviewed in full.
Choosing models
LangChain adds a managed browser to Deep Agents, with sign-in handoffs that keep passwords away from the model. Not reviewed in full.
Building agents
How LangChain tuned the browser tools its agents use, tracing inefficiencies and checking each change with evals. Not reviewed in full.
Building agents
Transluce trains models to predict how fine-tuning will change another model's answers, and finds scaling laws.
Summary in Issue #3Choosing modelsMarkets & safe use
What scientists actually use Claude for, from about 207,000 sessions and a survey of 528 researchers.
Summary in Issue #3AdvisingMarkets & safe use
Google's video model update: 15-second clips, conversational editing and lower cost. Not reviewed in full.
Choosing models
NVIDIA adds four open Cosmos 3 checkpoints for interactive world and driving simulation. Not reviewed in full.
Choosing models
The story behind Phonon, a small open text-to-speech model that runs on the device. Not reviewed in full.
Choosing models
LangChain's case for splitting an agent into an orchestrator that stays with the user and background subagents.
Summary in Issue #3Building agents
LangSmith can now keep traces and evals in a customer's own AWS, Azure or Google Cloud account. Not reviewed in full.
Building agents
A maturity framework for platforms that serve data products, across operating and investment dimensions. Not reviewed in full.
Advising
Mistral launches Mistral Large 4 and expands its Studio platform with managed agents and workflows. Not reviewed in full.
Choosing modelsAI ecosystem
An IBM explainer on whether vector databases are still needed for RAG and agents. Not reviewed in full.
Building agents
No Priors interview with Robinhood's CEO on tokenization and agentic trading. Not reviewed in full.
Markets & safe use
Three kinds of agent retrieval failure and how to fix each, using a competitive-analysis agent as the example.
Summary in Issue #3Building agents
Five techniques LangChain's Deep Agents use to keep long runs within the context window.
Building agents
A LangChain case study on Stripe's knowledge platform built on Deep Agents. Not reviewed in full.
Building agents
LangChain argues that companies should own their agents' data and memory to keep an advantage. Not reviewed in full.
AI business
No posts match. Try fewer words or another filter.
