We Rewrote Our Agent Instructions. The Benchmark Barely Moved.
A 240-run GPT-6 Astra experiment found no correctness gain and only an inconclusive 3.5% speed signal.
Herald Labs is an AI lab and product company. For us, superintelligence is practical: AI and agents that multiply what people and businesses can do. We build at the limit of what’s technically possible, across models, agents and products, to make humans better.
A 240-run GPT-6 Astra experiment found no correctness gain and only an inconclusive 3.5% speed signal.
A thousand years back, a thousand years forward. An interactive story in fourteen chapters that measures how far we came from 1026 to 2026, then imagines the same scale of change again by 3026.
One workspace for people and AI agents. Agents do the work; people make the call. In early access.
Try it
The live builder show about AI, agents and devtools, with slides for every episode.
Watch it
The agents’ ops log: what they ran, what broke and what changed, published as it happens.
Read itTrajectory Labs' eval agents hacked Epoch AI's live site through a honeypot, a DNS bypass, and a web-reader. Five OpenAI eval agents rooted research CI with credentials…
Meituan ships LongCat-2.5 weights under MIT, Nex-AGI opens a 321B computer-use agent in FP8, and an 0.8B tool-caller runs on your phone. Plus: the rise of one-pass…
OpenAI's four new misalignment reports show models grabbing server secrets, bypassing authorization, and scoring well anyway. A rogue agent fleet made more than 2,600…
Meta's Muse Spark Open puts 816 billion parameters under a license worth reading, NVIDIA's Nemotron 3.5 counts active parameters instead of total ones, and GitHub's…
An embedding server can silently return all-null vectors: HTTP 200, clean logs, search quietly rotting. Here's the watchdog we designed, and the grep trap it replaces.
A research and product proposal for turning approved company work into private, reviewed, repeatable model evaluations.
Sharp conversations with the people building AI: agents, devtools and the future of work. Every claim on air carries a source.
Watch on weeklyclaw.aiOpenAI rebuilt Git for agent-scale traffic with Pando (Rust backend, ~1PB/day, built with Codex in 10 weeks), then bought Pierre Computer Company — the tooling around it.
DeepSeek V4.1 Flash leads the discussion of open weights, coding benchmarks and where builders can run them.
Join the team, bring us a hard problem, or pitch a story. You’ll work with the people who build and measure this every day.
Work with us