Tooling for the outer loop
Latent Space published their write-up of the trends that defined AI Engineering at World's Fair 2026. The through-line: the field grew up. The interesting work is no longer the agent, it's the system around the agent. Harness engineering, evals, sandboxes, permissions. And the consensus shape of that system is a pair of loops: the agent runs the inner execution loop, and a human sits in the outer loop, setting direction and making the calls. One speaker compared the job to being a locomotive engineer, there to keep the thing on the rails.
We think the write-up quietly exposes an imbalance. Enormous engineering now goes into the inner loop. The outer loop, the human half of the arrangement, mostly still runs on memory.
The outer loop is conducted in meetings
Look at what the outer-loop human actually does: sets direction, reviews outcomes, decides when to intervene, carries the why. Now ask where those directions and whys come from. They get worked out in conversation. The roadmap call where scope was cut. The design review where an approach was rejected for a reason that took ten minutes to explain. The customer call that changed the priority.
The inner loop gets harnesses, feedback structures, and observability. The outer loop gets a two-line action item and whatever the human still remembers three weeks later. If the industry's new job description is "the person who steers," then the record of where you steered, and why, is load-bearing infrastructure. Almost nobody treats it that way.
Boswell is tooling for the outer loop. It records the meeting, transcribes it on your Mac, and files the decisions, the reasoning, and the action items as plain Markdown. The steering history stays legible, to you and to anything you choose to show it to.
Agents are just files, and so are your meetings
The write-up's fifth trend makes the fit concrete. Skills, the emerging standard for extending agents, turned out to be Markdown files. As one speaker put it, agents are just files; you write Markdown to give them capabilities. After years of frameworks, the industry converged on the plainest interface there is.
That's the format your meeting notes are already in, if you use Boswell. A skill file encodes procedure: how we deploy, how we review, how we write. A meeting note encodes judgment: what we chose, what we ruled out, what the customer actually said. When you drop last week's design-review note into an agent's context next to its skill files, you're doing the same engineering with better material. The skill tells the agent how work is done. The note tells it what this work is for.
The write-up also records a worry about skills: if everyone installs the same ones, everything converges and the output all looks the same. That's the strongest argument for the other file. Shared skills are, by definition, what everyone has. Your meeting archive is what only you have. The differentiated input to your agents was never going to be a downloaded best practice. It's the accumulated record of your own decisions, sitting on your own machine, in the one format every agent reads.
What we don't claim
Boswell is not a harness. It won't run your loops, gate your permissions, or eval your outputs, and the people building that infrastructure are doing necessary work. Our claim is one layer up, where the human sits. The outer loop's scarce resource is context about intent, and intent gets spoken. We keep it.
The industry spent 2026 engineering the system around the model. The system around the human is much older, and it still works: an accurate record of what was said, kept by someone who was in the room. The inner loop got its harness. The outer loop gets a biographer.