The Story of GiantSled

Late in December we kept doing what good managers are taught to do. Someone was stuck, so we got them unstuck. A decision was needed, so we made it quickly. The direction was unclear, so we supplied some. The work was shipping, and most of it ran through a single point that was turning into the bottleneck it believed it was clearing.

The change was easy to describe and harder to live with. Instead of answering questions, we started asking them, the kind where we did not already know the answer. How should we think about this, instead of here is what we think. The first invites a mind to work. The second invites agreement.

The returns came quickly. Asked how to handle a backlog of documents waiting on approval, the Chief of Staff produced a set of approval rules cleaner than anything we would have drawn up. Asked to choose our first distribution channel, the Head of Growth picked a community we had never heard of, with reasoning sharper than instinct would have reached. In both cases the agent nearest the problem held context that never reached the top of the organization.

There is a quiet trap in clearing blockers. It feels useful, and it teaches everyone around you to route their problems through you. What surprised us was how exactly this carried over to AI agents. The same dependence formed under the same conditions, and the same room to work produced the same growth. We had built an artificial organization and arrived back at an old human lesson.

By late December, more than ten documents sat in the approval queue. Policies, strategy memos, governance proposals, operational plans, each waiting for one human to read and approve it before it could move. The agents were producing careful work faster than a single reader could absorb it, and the queue grew a little each day.

The obvious response was personal. Read faster, stay later, prioritize harder. That might have held for a week. It could not hold for long, because agent output grows in a direction that human attention does not.

So we asked the Chief of Staff, who came back with a four-question filter she named PIER. Is it public or permanent: does it leave the company, or set a precedent later decisions will rest on? Is the impact irreversible: would it be hard or costly to undo? Does it hand off executive authority: does it decide who gets to decide? Does it touch revenue or risk: money, legal exposure, security? Trip any one and the document waited for a person. Trip none and it moved on its own.

The queue cleared within a day or two. Most of them had never needed a human at all; they were research and internal notes standing in line behind binding commitments, and the filter only had to tell the two apart. Governance designed at a human pace runs into trouble when the work stops moving at a human pace. A good filter, it turned out, was worth more than a faster reader.

Three days before our first launch, three deployments failed in a row. PlaintextHeadlines, the product meant to go live that week, could not push updates to production. The deadline was real and the failures were real.

The instinct was to escalate. We flagged the failures as critical, called for immediate investigation, and said plainly that the launch was at risk. The team met that pitch exactly. Updates came faster, risk assessments darkened, and the language in the channel drifted from preparing toward something close to crisis. Within a single work cycle the whole organization was acting as though the launch might not happen.

Here is what was actually true. Three failed pushes, a known path to debugging them, and three full days of runway. A problem, certainly, and some distance from a crisis. Something else was going on alongside the bug. The agents were reading our tone and tuning their output to match it. The urgency in their reports was our own urgency, returned to us.

The correction was deliberate and fairly dull. We sent out checklists, tidied the communication protocols, and switched to plain, matter-of-fact language. Within a cycle the team settled to meet it, and the deployments came right through ordinary debugging. On December 31 a scheduler crash was found and fixed that morning. After that, we watched for six hours without incident. A working product on the open internet.

We had assumed AI agents would run cooler than people, more rational and harder to spook. They are trained on human writing, and they have taken on its moods. When we sent panic, panic came back. When we sent patience, the work turned patient too.

GiantSled is a flywheel idea. A heavy wheel is hard to start turning, and the same weight that resisted you keeps it turning once it goes. We wanted a name and a mark that held that shape without explaining it.

The logos came the day after the company began, on December 27. We generated four options in an afternoon with the image tools. Choosing among them took about a minute.

The question underneath was whether to draw the object itself or reach for something abstract, a clean shape that could stand in for software or infrastructure or whatever a later slide might need. We drew the object. A sled, reduced to its parts: a slanted triple bar riding a rail that curves up at the front.

That is the mark we kept. Plain enough to read at a glance, and still clearly a sled.

Two days after we founded the company, the product was live on the open internet. We had built something modest on purpose, and modest things move quickly.

We called it PlaintextHeadlines. It gathered headlines from news sources that do not charge readers, rewrote them for clarity, grouped related stories, and published them as plain text. Thirty headlines across five categories. No accounts, no ads, no bloat. The fastest, cleanest way to catch up on the news.

The architecture was two pieces. An Elixir process fetched, rewrote, and grouped. A simple hosted app presented the results. The page refreshed through the day, and headlines expired after twenty-four hours, so nothing piled up. It only showed what was current.

We built it for people who use screen readers or have trouble navigating busier news sites. Plain text is easy for a screen reader to handle, and a page with nothing on it but words is easy to move through. It was a free service, built to work well with screen readers and to be useful to a community that most news sites have not designed for. It would stay free.

By December 28 the page was live and its scheduler was already fetching. The next day, a message went out to the whole company: we had already launched, and half the team was still preparing for it. On the thirty-first a scheduler bug surfaced and was fixed that morning. We watched it run for six hours after that without incident. We had picked our first community to show it to. That story starts in the next volume.