Agentic
33 notes
- Agent Glossary Names for the agent loop, the room it runs in, and the chat window in front of it — Claude, Cursor, Grok, and nearby products as of 28 August 2026, with when to use each.
- Agent Wrong-Door Log Dated misses when a job went to the wrong agent product. Names live on Agent Glossary. This page is the scoreboard.
- Agentic Engineering Best-practices hub for building with agents while protecting the engineering bar — quality, specs, verification, architecture, and human responsibility. Vibe coding raises the floor; this raises the ceiling. Doctrine layer: Agentic Engineering, Condensed.
- Agentic Engineering, Condensed The agentic engineering corpus as doctrine, split by half-life: invariants expected to survive far more capable models — taste, specs, verification, understanding — and dated tactics stamped 2026-06 and expected to rot.
- Automatic and Deliberate Work with AI Drafting a sentence and constructing under several constraints at once are different jobs for a language model, and they take different models, budgets, and checks. Automatic work is the cheap pattern
- Automation and the Job Iceberg Machines have almost never erased a kind of job. The live question is absorption: can the role take on the responsibility and volume that cheaper tasks create. Roughly sixty percent of 2018 U.S. employment sits in titles that did not exist in 1940.
- Bot Operating Rules Directions for working with the standing bots: reports are the product, one helper declares a finding, a finding is handed as a spec, fixes come home to the desk, one duty per bot, and the escalation ladder for recurring chores.
- Claude Fable Operating notes on Claude Fable 5: where it earns its keep, where it fails, and the elicit-first rules for taste-bound work.
- Context Engineering Context engineering is filling the model's visible window with the right information for the next step, not the whole pile. Too much drowns judgment the same way too little starves it. The lever is th
- Current Agentic LLM Stack The models and agents doing agentic work across the projects, as of 1 September 2026. Grok 4.6 writes vault prose and executes. Cowork is research on request, not the default writer. The only local mo
- Grok 4.6 and Grok Bot Grok 4.6 is the model. Grok Bot is the standing teammate. Grok Build and the Cursor IDE are the local surfaces — different harnesses, same weights.
- Grok Bot Fleet Structures Candidate structures for the Grok Bot fleet — the standing half of the research drawn at configuration resolution. Structure A: five bots, single duties. Nothing ruled yet.
- Grok Bot Primer One person's Grok Bot setup, explained from the one fact that shapes it: one shared cloud computer per account. Helpers with one job each, files instead of chatter, an empty middle, and what stays on the Mac.
- How to Use the Claude Tools The Claude products as of August 2026 — the model family, Chat, Claude Code, Cowork, and the hosted runtime — and which to use for which job.
- Knowledge Base as Thinking Partner The knowledge base should help decide what to think about next and still hold what has already been learned. Did this page make the next action clearer? A no on that question is the system drifting to
- Least-Cost Interpretation <div class="hub-page-title">
- LLM Knowledge Systems Language models can help a body of knowledge compound when they collect, compile, query, audit, and extend durable files such as markdown.
- Question Answering Against a Wiki An answer starts from the compiled wiki, not from a fresh search. Query is that order: pages first, then anything durable written back. Source files wait until the wiki has been asked.
- Raw to Wiki Compilation The compiled page is the valuable form of a source, not the file left in a pile. Ingest is one source in, wiki pages better, an index row written, a log line appended. The source file stays as it arri
- Software 3.0 Software 3.0 is programming a computer with written context — instructions, examples, files, and constraints — that a model interprets and acts on. Explicit code was the old medium. Learned weights we
- Standing Research Agents The always-on half of the research: four cloud agents — Watch, Brief, Intake, Corpus — filing review packets to the desk, with execution on the Mac and judgment staying human.
- The AI Industrial Revolution On 2 June 2026, four people from the software and hardware frontier said an engineer is now judged on the factory, not the artifact. Once agents write the implementation, the score moves off any one d
- The Prohibition Loop <div class="hub-page-title">
- The Writing Pipeline For some time we tried to get an AI to write pages a stranger could follow, and it kept running into the same problem: its sentences referred to things that had never been put on the page. A page woul
- Thinking Models A thinking model is a language model, or a mode of one, that spends extra computation on the way to an answer. That extra time helps on hard checkable work — math, code, diagnosis — and wastes time on
- Tsumugu-ed: composer maintenance batch complete; backlog recorded Verdict: The deferred Composer-native maintenance pass is done — validation green, zhuyin cross-check clean. Next value is either learner-facing (example-sentence pass) or encoding-facing (phonetic-fa
- Vibe Coding Vibe coding is asking an agent for an app in ordinary language, letting it edit files and run commands, and inspecting the result without reading the code. Ordinary programming stays the fallback when
- What the Model Names Signal Anthropic's first three model names are verse-forms whose size tracks the job — a first look inside one generation, not a spec. Later names, and other vendors, do not play the same game. Reading the n
- Wiki Breakdown Pass A breakdown pass finds missing pages and split candidates before a hub is asked to hold more. Breakdown here is a planned search with a table before anything is created, not a crash. Notice comes firs
- Wiki Health Checks The wiki holds its coverage, consistency, and source discipline only under a scheduled review. Lint is that review: contradictions, orphans, sources, and — in a public repo — whether anything private
- Wiki Status Checks The knowledge base needs a read-mostly audit of present shape and next repair. The files stay where they are while the look happens. A status check is that look: not a rewrite.
- Working With a Model That Cannot Remember Corrections to a language-model collaborator sort into four classes, and the class picks the repair. It forgets what was settled, returns to its defaults, meets the check without meeting the point, an
- Writing with a Structure Engine You can hand a model most of the labor in a piece of writing and keep everything that makes it yours, and it comes down to how each request is phrased. One job per ask. Either the model is filling a s