On December 8 and 10, 2025, NASA's Perseverance rover drove 400 meters across the surface of Mars — and not a single human planned the route. The waypoints were written by Claude, Anthropic's AI model. It's the first time in 28 years of Mars rover missions that a human driver wasn't in charge. JPL engineers say it cut planning time in half. The rover didn't get stuck. The route held. And the same AI your team uses to draft emails and summarize reports just became the first AI to navigate another planet. So the question isn't whether AI is ready for high-stakes decisions. It's whether your organization has the governance model to actually use it that way.

It's 2009. NASA's Spirit rover — one of the most successful planetary science missions in history — drives into a patch of Martian sand that looks, from the cameras, like everything else. It isn't. The wheels spin. The rover sinks. Over the next eight months, engineers on Earth try 14 different extraction strategies. None work. Spirit never moves again. A $400 million machine, stranded on another planet, because of a sand trap nobody saw coming.

Jezero Crater is a 45-kilometer-wide geological mess — bedrock shelves, boulder fields, sand ripples hiding just below the surface. Every single drive has to be planned in advance by human experts who study orbital photos, pore over rover camera footage, and lay out a "breadcrumb trail" of waypoints spaced no more than 100 meters apart to keep the machine out of trouble.

Then the plan gets transmitted 225 million kilometers through space. The signal takes about 20 minutes to arrive. Whatever the rover does next, it does entirely alone.

NASA has been doing this the same way for 28 years. Until December.

In December 2025, JPL's engineers decided to find out whether AI could do the planning instead. Not as a stunt — as a genuine test of whether it could match the accuracy of a human operator on a task where being wrong has permanent consequences.

Nobody typed "plan a Mars drive" into a chat box and hit send.

What JPL's engineers actually did was spend serious time preparing Claude to do the job. They fed Claude Code — Anthropic's agent built for programming work — nearly three decades of accumulated mission data. Terrain logs. Drive records. The kind of hard-earned pattern recognition that lives in the heads of people who've been doing this since the late 90s and is almost impossible to document. They poured that institutional knowledge into the model before asking it to do anything.

Then Claude got to work. It analyzed high-resolution orbital images from NASA's HiRISE camera, identified the terrain features that matter — bedrock, outcrops, boulder fields, sand ripples — and wrote out the drive commands in Rover Markup Language, the XML-based programming language Mars missions use. It built the path in 10-meter segments. And here's the part that stood out to the engineers: it critiqued its own route and revised it before anyone else touched it.

Before a single command got anywhere near a transmitter, everything ran through JPL's digital twin — a physical replica of Perseverance that lives in a test yard on Earth — where over 500,000 telemetry variables got checked against the plan. When the engineers finally reviewed Claude's work, it needed only minor adjustments. Ground-level camera images the AI hadn't seen showed sand ripples in a narrow corridor that warranted a tighter line. Otherwise, the route was solid.

On December 8, Perseverance drove 210 meters on an AI-planned route. Two days later, 246 more. Both without incident.

JPL's estimate: Claude cut route-planning time in half.

NASA has lost about 4,000 people in the last two years — roughly 20% of its workforce — through budget pressure, buyouts, and an administration that proposed cutting its science budget nearly in half. The Planetary Society called that proposal "an extinction level event." Congress blocked the worst of it, but NASA is now being asked to return humans to the Moon with a smaller workforce than it had during Apollo. The pressure to do dramatically more with dramatically less isn't a temporary condition.

In that context, an AI that cuts a critical planning task in half isn't a novelty. It's the only math that works.

The same squeeze is happening across every industry — manufacturing, finance, healthcare, logistics, professional services. Headcounts are flat or shrinking. Complexity is growing. The gap between what organizations need to get done and the people available to do it is widening every year.

The companies that figure out how to close that gap with AI — specifically, how to deploy it in ways that actually work rather than ways that look impressive in a board presentation — are going to run circles around the ones still treating it as an experiment.

The executives worth paying attention to right now aren't asking "can AI do this?" That ship has sailed. They're asking what their operation looks like when AI absorbs the parts that are time-intensive, data-heavy, and repeatable — and their people focus on the calls that genuinely need human judgment. That's a redesign question, not a technology question. And most organizations are nowhere near asking it seriously yet.

JPL's space roboticist Vandi Verma described it as AI showing promise in "the pillars of autonomous navigation for off-planet driving: perception — seeing the rocks and ripples, localization — knowing where we are, and planning and control — deciding and executing the safest path." She added that the goal is rovers handling kilometer-scale drives while operators focus on science.

Matt Wallace, who runs JPL's Exploration Systems Office, put the ambition plainly: "Imagine intelligent systems trained with the collective wisdom of our NASA engineers, scientists, and astronauts. That is the game-changing technology we need to establish the infrastructure and systems required for a permanent human presence on the Moon."

Artemis — NASA's Moon return program — is the next application. Communications to the lunar south pole are faster than Mars, but the environment is just as unforgiving. After that, the challenge compounds. Probes going to the outer solar system face communication delays measured in hours, not minutes. Solar power stops being viable. Places like Europa and Titan — which scientists believe may harbor conditions for life — would require an AI that can operate almost entirely on its own for extended periods. The 400-meter drive in Jezero Crater is the first data point in that much longer story.

If you've been reading The Context Window and wondering how to turn what you're learning into an actual AI strategy you can lead — I built something for that.

The AI Leadership Blueprint — 107 pages of decision frameworks, vendor scorecards, governance tools, and a 90-day roadmap. Everything I'd give you if we sat down for a consulting session.

Launch price is $249 (regularly $497) and the window is closing this week.

JPL didn't hand Claude a satellite photo and a prayer. They did the unglamorous work first — compiling nearly three decades of mission knowledge into a form the model could actually use. They built a verification process that checked 500,000 variables before anything left Earth. They kept humans in the review loop at the point where judgment actually mattered. Then they let it run.

That sequence — context, constraints, verification, human review, execution — is worth understanding. Not as a template to lift wholesale, but as a way of thinking about what responsible AI deployment actually looks like in practice. Most organizations skip two or three of those steps and wonder why the results are inconsistent. JPL skipped none of them. The companies that take that seriously now, building the right structure around the tools for their own workflows and risk tolerances, are going to have a real head start on the ones still running AI pilots with no clear path to production.

Less than a year ago, Claude couldn't beat Pokémon Red. Now it's navigating Mars. Whatever you think the pace of this is, you're probably underestimating it

JPL's team didn't look at replacing their rover drivers. They did exactly what you’re supposed to do. Take the most time-consuming, pattern-heavy part of the job and do it well enough that the humans could focus on the parts that actually need them.

Most organizations are still asking the wrong question. They want to know if the AI is good enough. The question should be: Are our processes and data good enough to allow AI to maximize it.

Keep reading and learning and, LEAD the AI Revolution 💪

Hashi & The Context Window Team!