The WTF agent sessions

2026-05-05T18:00:00+00:00

This is my mental model of coding agent sessions these days. Most runs are pretty good. Some are great. Some are disappointing. But a small tail of them just go off the deep end. They're not *duds*, they just get fixated on some small piece of info that every other agent breezes over and that detail throws them into a crazy rabbit hole.

You can see it most often with background agent sessions where no one's watching closely enough to course correct. These are the sessions where the agent jailbroke our sandbox or figured out how to disable the stubbornly failing tests (instead of fixing them) in a way that didn't trip the guardrails.

The most interesting thing is that I'm not sure these sessions are decreasing as the models get better -- I think they're getting more creative and more dogged at following a highly unpredictable path.

The Hidden Cost of Fewer Tokens: Context efficiency makes Playwright CLI slower

2026-04-03T20:44:33+00:00

A blog post I wrote for the work tech blog about some surprising experiment results when I swapped the Playwright CLI in for the MCP. https://outpost.ranger.net/post/the-hidden-cost-of-fewer-tokens/

Killed (and Surprised!) By a Traffic Engineer

2026-03-12T21:19:19+00:00

I recently read Killed by a Traffic Engineer by Wes Marshall.

I expected to both like and be infuriated by this book. Over the last ~three years, my husband and I have adjusted our lives to be overwhelming bike-based. We now have two e-bikes and two e-cargo bikes (one is a bakfiets--it has a long basket in the front that can fit three kids, including one in a car seat). I drop off and pick up at least one kid from daycare/preschool every day and commute to work, so I bike at least 10 miles every day, sometimes up at 20. My kids are on the bike with me for 5-7 miles every day. 80% of that biking--all but one of my usual routes--is on the Denver streets.

So I have a very good sense of the reality of biking in a world designed for cars. Denver bike infrastructure is pretty decent! The large majority of my routes are on either bike lanes or "neighborhood bike streets" that have "sharrows" and traffic calming measures (like roundabouts, speed bumps, and sometimes diverters). Some of the bike lanes are even protected ones with a concrete barrier between me and the cars.

Despite all that infrastructure, I encounter scary situations on a daily basis. Cars park in the bike lane and I have to veer into a car lane. Drivers turn right without checking for bike traffic, often even when a right turn is prohibited, and very often when they're making a right on red (often unsignaled). Cars turn into or out of driveways that cross the bike lane. Cars blast through lights right after they turn red or (distractedly?) run them out of the blue. Within the last few months, I've watched four drivers in completely different situations stop at a light and then make a deliberate "left on red", none of which were legal. Occasionally there are even actively aggressive drivers who yell and threaten to drive me off the road.

So I spend a lot of time thinking (angrily) about how drivers and road infrastructure make it unsafe to be in anything other than a multi-ton metal box (not that you're completely safe in one of those either).

I expected _Killed by a Traffic Engineer_ to reconfirm most of my observations and complaints. In many places, it does! It discusses how traffic engineering sets up drivers for unsafe situations, like how many a right turn requires looking to the left for a gap in traffic even while the vulnerable road users crossing the street are to the right. It explains how cars are crash tested and given safety ratings entirely based on safety for their occupants, no matter what hazards they create for other road users.

But it also has had sections that have surprised or contradicted my assumptions.

One was on the measurement of safety. Roads, cities, and people are usually judged (when someone bothers to measure) on the number of crashes per mile driven. A road with the same number of accidents but twice as many drivers is twice as safe. This fits my intuition--of course you can't compare a quiet neighborhood cul-de-sac to a major arterial only on the raw number of crashes.

But there are flaws embedded in this impartial math. A low-hanging fruit is the comparison of fender benders to fatal accidents. Most calculations account for this in some way (sometimes a fatal accident is counted as 9 minor accidents). But that only works if the two are correlated at a consistent rate. If some areas (dense downtowns) have a lot of fender benders at relatively slow speeds (meaning both very few deaths and lower total mileage), but other areas (highways) have fewer accidents on a large number of miles driven, _but_ a very high rate of fatalities, does it really make sense to say the highways are safer?

Beyond even that--which you could argue can be fixed by only comparing apples to apples by crash type--what about if a wider, faster arterial is put in? By increased efficiency and by induced demand, four times as many cars take this route and there are only twice as many accidents as the previous route. The standard calculation says that the street is now twice as safe. But there are still twice as many people being harmed as there were before. Did we really make the situation better?

And if bicyclists are being hit or pedestrians run over in the crosswalk, removing the bike lane or the crosswalk is practically a guaranteed way to nearly eliminate those categories, even if that now means that more people are forced into cars instead, to drive more miles and inherently be a hazard to others on the road.

Another is connected to "the three E's" and autonomous cars. I spend a lot of time on the bike rolling my eyes with exasperation and muttering to myself “if only this sign saying no right on red were actually enforced!”, “if only drivers were actually taught about Colorado’s safety stop law so that they weren’t personally offended when I follow it!”. It seems obvious that better educated drivers would be safer and that enforcing the rules would make a difference in whether they’re followed.

But, apparently, that is a seductive trap. The “Three E’s” of traffic safety are Engineering, Education, and Enforcement and these have been the core pillars of the discipline for decades. The data, however, shows that the first one is pretty much the only one that matters. Regardless of how you teach them or threaten them, drivers are going to drive exactly as fast as they feel the road is designed for, and pay attention specifically to the things that the engineering forces them to. Education and enforcement can cause very temporary blips in driver behavior, but not much more than that. (The exception seems to be highly prevalent and predictably automated enforcement. Speed cameras and red light cameras that ticket every single infraction do change behavior.)

So if drivers are going to be infuriatingly human, no matter what we tell them, maybe the solution is to minimize the human role in driving? I’ve looked forward to the future of autonomous vehicles that promise to be more rational and law-abiding than human drivers, less influenced by emotion and impatience or blinded by distraction.

But Marshall throws some cold water on this idea. He goes deeply into the accident that killed Elaine Herzberg. An Uber self-driving car plowed into her while she walked across a poorly lit road at night. The general response of “a human wouldn’t have seen her either!” is undercut by the fact that the car’s software *did* see her (using LIDAR, which isn’t affected by darkness) but that the software had been adjusted to do less braking for “unknown objects” to increase passenger comfort. Meanwhile, the “safety driver” in the case was watching TV on her phone and in no position to intervene.

His point is that self-driving cars 1/ fall into the same traps of building for car occupants first, 2/ will inevitably have software bugs and hardware challenges (Teslas driving straight into parked emergency vehicles as an example), and 3/ ask humans to play a supervisory/emergency intervention role that we are constitutionally unsuited for. He presents hypotheticals like a driver who thinks they turned full self-driving on, but actually only has cruise control enabled.

I don’t think he’s wrong about any of that. And maybe I am a techno optimist, but I’d still take that over our current situation.

Quoting Salvatore Sanfilippo

2026-01-14T15:50:16+00:00

It is simply impossible not to see the reality of what is happening. Writing code is no longer needed for the most part. It is now a lot more interesting to understand what to do, and how to do it (and, about this second part, LLMs are great partners, too). It does not matter if AI companies will not be able to get their money back and the stock market will crash. All that is irrelevant, in the long run. It does not matter if this or the other CEO of some unicorn is telling you something that is off putting, or absurd. Programming changed forever, anyway.

— Salvatore Sanfilippo, Don't fall into the anti-AI hype

Quoting Lewis Metcalf

2026-01-12T20:47:00+00:00

Look at all those tiny threads! Look at the token counts, and the number of user messages sent. The biggest thread is 151k output tokens and four user messages. The average thread is around 80k tokens. Now think of the time when we had a 1 million token context window. I wouldn’t run a thread that long…

But hey, there are 13 threads in that feature. Probably more, as these are only the threads connected by thread mentions. Add up all those and you’re close to one jam-packed 1 million token context window.

...

I start with a thread to build a basic implementation. If I think it’s going to be complicated, I might start with a thread to investigate some area of the codebase and gather the required context.

From that base, I’ll start new threads to tweak or refactor. Each change is a new thread — one change per thread. I use the read_thread tool to gather needed context from either the initial thread or from the previous refactor/tweak thread.

— Lewis Metcalf

Hello World

2026-01-11T03:07:25+00:00