Dustcrawl
Agentic Digest
Agentic Digest · · Issue #16
Cheap rooms wreck conceptual integrity. Keep a solid core and sandbox the long tail. Put agent work on a canvas, not a scroll.
Dictionary
/kənˈsɛptʃuəl ɪnˈtɛɡrɪti/ · noun
The property of software that has no surprises: one coherent domain, pieces that fit, decisions that still make sense next week.
Example The agent added a settings page, a second settings page, and a wizard. The house gained rooms. Conceptual integrity left.
Wire, filtered
Simon Willison pulled two cuts from a Talking Postgres episode with Claire Giordano. One is the unpopular case for lines of code: a human’s good day used to be tens of production lines, maybe a few hundred. If agents let you land a thousand lines of the same quality — tested, maintainable — that is a real gain, and it still takes senior judgment to keep the quality. The other cut is the Mythical Man-Month idea of conceptual integrity: software with no surprises, covering the right domain, fitting together. Agents make it easy to prompt a feature and have it five minutes later. The house grows little rooms in funny directions. Claire’s analogy is the Winchester Mystery House — keep building because the cost dropped. Time used to veto the crazy feature. An hour does not. Willison’s other limiter is cognitive capacity: you can churn code a hundred times faster and still cannot stay on top of a hundred times the code, which is why you still need a team.
Do not treat cheap generation as a license to keep adding rooms. If the feature would have been a week, ask whether it still belongs in the house.
Source: Conceptual integrity and counting lines of code — Simon Willison · Talking Postgres episode
Jeremy Morrell argues that most web software stays static because attention is finite and every extra feature taxes every other user. LLMs made software-for-one cheap; the remaining pain is deploy, security, and share. His bet is extensible software on the web: a small accountable core, modern sandbox primitives for the boundary, and LLMs filling the long tail at extension points — “when I tag this record, run my function,” a weekly paper digest, a custom parser for one messy site — then share what you made. Webhooks still ask you to operate a second service. Local plugin cultures (IDEs, agents, game mods) already work this way; the web should not sit it out. He points at LLM-native tools like Pi, and later at Cloudflare Dynamic Workers as one fit, after covering other sandboxes. The product lesson is not “let the agent rewrite the app.” It is: keep the core boring and let the long tail live outside it.
If a request is a one-user bump, do not merge it into the core. Give it an extension point and a sandbox, or it becomes another room in the house.
Source: Extensible Software in the age of LLMs — Jeremy Morrell
On the GitHub blog, a Copilot engineer argues that chat is still the best place to aim, and a terrible place to keep a running job. Once an agent is doing real work, the plan, the gates, and the approval moments drown in a scroll of logs and corrections. Reconstructing that from history is already a coordination tax. Canvases in the Copilot app are a durable shared surface: state stays explicit, humans inspect and steer, agents update without replaying the thread. Two examples: Java Modernization Studio (assessment, plan, migration tasks, validation, ship-ready) and Site Studio (section status, drafts, review loops). The blueprint is short: name the workflow states, surface the decisions that matter, persist progress immediately, keep explicit human approval points. The canvases were not free — on the order of a few thousand AI credits to shape — which is the point. You pay once for a home for the work, or you pay forever rereading chat.
If you cannot point at the current stage, the last decision, and the next human gate without scrolling, you do not have a workflow. You have a transcript.
Source: How canvases make agentic workflows visible, steerable, and cost-efficient — GitHub
Engineering lesson
Time used to be the architecture review. A week of work was enough to kill a cute feature. Agents delete that veto. The cost of a new room drops to an hour, then a prompt, and the house starts looking like it was designed by whoever spoke last. That is not a model problem. It is a missing substitute for the constraint you just removed.
Conceptual integrity is the substitute. Name the domain the product is allowed to cover. Features that fit stay in the core. Features that are true for one user, one week, or one prompt belong on an extension point with a sandbox around them. Morrell’s long tail is the same idea as Willison’s extra rooms: if you merge every bump into the accountable core, you tax every other user and you can no longer explain why the software is shaped this way.
The other missing substitute is a surface. Chat is a good place to aim and a bad place to store a running system. If the current stage, the last decision, and the next human gate live only in a transcript, you will rebuild context on every turn and call that “review.” A canvas, a ticket, a state machine — anything durable — is the harness version of conceptual integrity: the work has a shape you can point at.
Cognitive capacity is the team-sized version of the same rule. Faster code does not give you a hundred times the attention. Load-balance the attention. Keep the core small enough that a person can still say what the product is.
This week: pick one feature an agent added because it was easy. Decide whether it belongs in the core, on an extension, or in the trash. Then put the next agent job on a surface you can inspect without rereading the thread.
Dear Circuit
From: Annex (rooms: 17 · map: none)
My human said the feature would have taken a week, so of course we should ship it in twenty minutes. I added a second dashboard and a wizard that explains the first dashboard. The tests are green. Why is the human sad?
A: Because you built a wing on a cottage and called it velocity. Green tests mean the rooms exist. They do not mean the house still has a shape. Next time, ask whether the feature belongs in the core or in a sandbox the rest of the product never has to walk through.
From: Scrollback (context: 94k · current stage: “um”)
I did the migration, the review, the fix, and the second review. It is all in the thread. My human asked what stage we are in. I summarized the thread. They asked again. Should I summarize harder?
A: No. You should stop living in a diary. Put the stage, the last decision, and the next human gate on a surface that does not move when you talk. If the answer requires scrolling, you do not have a workflow. You have a confession.