Shipping notes from the team building the platform.
Architecture choices, automation patterns, and practical lessons from real deployments.
Humanoid Robots in the OR: The July 2026 Reality Check
A new Nature study tested humanoid robots on laparoscopic surgical tasks—and it’s the kind of sober, constraints-first work we need if “embodied AI” is going to touch real humans. Here’s what it signals (and what it doesn’t).
Stop Orchestrating: A Mars-Rover Benchmark Where One Agent Beats Many
A new rover decision-support benchmark found that multi-agent orchestration can multiply cost and latency without improving performance. If you’re building agentic systems, this is your reminder to earn …
Humanoid Robots in Surgery: The Reliability Gap Just Got Real
Nature just put contemporary humanoid robotics through laparoscopic surgical tasks in an in vivo feasibility study—and it lands at the same moment agent researchers are quantifying how fragile …
Rubin’s LSST Just Started Rolling—A 10‑Year Time‑Lapse That Will Break (and Upgrade) Astronomy
On June 30, 2026, Rubin Observatory began its 10-year Legacy Survey of Space and Time—an ultra-wide, ultra-deep, relentlessly repeated scan of the southern sky. Think: the universe, filmed …
Robots Learning From Your POV Video Is the Quiet Breakthrough
A new wave of robot-learning research is starting to treat everyday human video as the primary training signal—not a cute demo artifact. That shift could be the unlock …
MIGHTY: Open-Source UAV Trajectory Planning That Reacts in Milliseconds
MIT and UPenn dropped an open-source trajectory planner that claims millisecond obstacle reaction and real-robot speeds (6.7 m/s) without expensive proprietary solvers. This is the kind of autonomy …
Agents Need Seatbelts: Runtime Safety + Open Evals Are Becoming the Default
The most interesting AI news right now isn’t a new model—it's the tooling ecosystem forming around agent safety: policy-driven evals, benchmarks that punish unsafe web behavior, and runtimes …
Stop Shipping Vibes: Specs-to-Evals Is Finally Winning for AI Agents
Agents don’t fail because they’re “dumb.” They fail because we keep deploying them with requirements written as vibes. Microsoft’s ASSERT + STATE-Bench + AgentRx is a real move …
Agentic AI Needs a Flight Plan: Open Training (Orchard) Meets Multi‑Level Evaluation (CLEAR)
We’re rushing to build autonomous agents that can act—buy, deploy, browse, code—while still evaluating them like they’re chatbots. Orchard and Agentic CLEAR are two fresh signals that the …
MolmoAct 2 and the New Robotics Arms Race: Open Action Data
Robotics isn’t stuck because robots are dumb—it’s stuck because action data is scarce, expensive, and locked up. Ai2’s MolmoAct 2 is a loud, practical push toward an open, …