The Pulse, Friday 4 September 2026
OpenAI shipped GPT-6 Astra, its first model rated Critical for cybersecurity capability, with strong eval gains across coding and reasoning. Anthropic opened Claude Code to custom tool extensions and Nous Research shipped one-click local models in Hermes Desktop. Yesterday's simultaneous ChatGPT, Claude, and Grok outages are still unexplained.
Models
OpenAI: GPT-6 Astra
OpenAI's new flagship: recurrent architecture for cheaper long tasks, first model rated Critical for cybersecurity capability (vendor claim).
Our agent stacks run on models like this via OpenRouter; the Critical rating also means review discipline matters more on anything it touches.
Google DeepMind: WeatherNext 3
DeepMind's most accurate global weather model yet, in production with partners.
No direct use for us yet
Research
DuplexSpeechBench-IFEval: Evaluating Implicit Instruction Following in Speech
Benchmark for whether voice agents follow instructions that are implied rather than stated, the case our Vapi flows live in.
Directly relevant to EverRaise voice agents: implicit instruction following is where call flows break.
When Users Don't Ask: Benchmarking Context-Driven Memory Retrieval in Conversation
Benchmarks retrieving the right memory when the user never explicitly asks, keyed off conversation context.
Applies to our CRM and sales agents: surfacing the right lead context without being prompted is the whole game.
RuleMem: Active Rule Memory for Long-Term Conversational Agents
A memory layer that keeps standing rules active across long conversations instead of letting them decay into context.
Maps to how we keep agent behaviour stable across multi-week client deployments.
SWE-Gate: Passing Functional Tests Is Not Enough for Software Engineering Agents
Shows coding agents that pass test suites still fail real task requirements; proposes a harder gate.
Calibrates how much we trust green tests from delegated agent work, which is most of our review burden.
Releases
LangChain 1.4.0
Stable LangChain release with graph and streaming refinements.
We build client agents on LangChain; worth bumping on the next touch of each service.
Langfuse v4.28.1
Patch release for the v4 line of our tracing stack.
We run Langfuse for agent observability; patches get picked up on deploy.
Hacker News
Ask HN: Why were OpenAI, Claude, and Grok simultaneously down?
All three major providers went down in the same window yesterday; no shared root cause confirmed yet.
Our multi-provider routing via OpenRouter is exactly the hedge for this; a reminder to keep provider fallbacks configured.
Which tools do Claude, Codex and Cursor choose? We measured 17k runs to find out
Measured which CLI tools coding agents actually install across 17k runs.
Direct input for how we scope delegated agent work and what tool access matters.
Porting my 1993 Amiga game to Godot, with an LLM reading the 68000 assembly
A real-world account of using an LLM to port 68000 assembly to Godot, with honest notes on where it needed human correction.
A good calibration read for our own agent-delegation boundaries.
Qwen 3.8 27B available on Cerebras at 1500 tokens/s
Qwen's new 27B open model served on Cerebras wafer-scale hardware at 1500 tokens per second.
Open-weight models at that speed change what we can run client-side; worth a benchmark pass against our API defaults.
People on X
François Chollet: frontier models vs ARC-AGI-3
Chollet notes frontier models still score near zero on ARC-AGI-3, released in March, and puts today's capability leaps in that context.
The calibration counterweight to launch-week claims: ARC's author on where the frontier actually stands.
ClaudeDevs: extending Claude Code with custom tools
Anthropic is exploring a new way to extend and customize Claude Code with custom tools.
Our whole delegation workflow runs through Claude Code; custom tools could replace several glue scripts.
Nous Research: Hermes Desktop one-click local models
Hermes Desktop now sets up local models in one click, reading config and wiring the runtime automatically.
The box we run on is Hermes-first; local inference is now a one-liner for client demos without API spend.
ClaudeDevs: `ant apply` for Managed Agents
New `ant apply` command declares Claude Managed Agents declaratively from the CLI.
Declarative agent config is the shape we want for repeatable client setups.
Python, Node, Next, DSPy, OpenRouter models, Hugging Face orgs, and Z.ai had nothing new in the window.