AI

The interactive session that runs everything else. Opus for heavy fan-outs, Sonnet for cheap classifiers, Fable when available. Fable was the first model I would treat as a taste authority.

The workhorse under every AI script, routed through one provider. Also the voice mode, which turned language practice into arguing philosophy in French on the elliptical.

The same problem gets fed to two models and the outputs compared.

It handles anything over a hundred thousand lines in the summary pipeline.

One always-on session rooted in my personal system that reaches into every other repo. The best tool I have for thinking in code.

Writes the code and bridges the chat subscription into an SDK provider. I flip between it and Claude Code based on which has usage left.

The agentic browser. Anything that has to happen on the web goes here. It is code-mode rather than tool-mode: the agent writes programs against a REPL with browser bindings, so twenty actions cost one round-trip. It is also a sensor that reads without raising; it checked a jury-duty page daily and never surfaced the date.

They are dispatched off Linear issues when the other two run out of usage.

Sixteen always-on agents on the Linux box, one per business. Good for asking one body a question; the fleet view is the gap I am building for.

A repository defines an agent and the agent lives on its own machine as a persistent process. I run a fleet on it to learn what my own runtime needs to be.

Five months as the daily interface. Retired as an interface in August; it survives as the subprocess that runs test fleets.

It solved provider abstraction and did not pretend to solve the rest.

Replaced typed journals within a month. A model turns six hours of rambling into a transcript with a spine.

The walking recorder took over from it in April.

A hundred and thirty meetings, a search collection of their own. Coding sessions with an agent are meetings too.

I drop in the reading and get a podcast for the gym.

Keyword and vector search over eleven years of journals, chats, and transcripts, reranked. Content production becomes retrieval, not generation.

Benchmarked against the fast model across twenty queries; kept because losing it noticeably degrades search.

Facts with a valid-from and a valid-to, distinct from text search.

Launching my own AI commands from a keystroke was the first time a desk felt like a cyborg setup, in 2024.

Affirmations and concept rehearsal in my own voice. Last used in April 2026; still wired up.

Fan out twenty videos, mark the bad ones, and the next batch is better.

I never found the happy path in it; the concepts kept multiplying faster than my understanding of what I needed.

I evaluated it twice. It is built for workflows, and I want autonomy.

I used it for four days. It wants a separate agent per thread and connects only through approved plugins, and I wanted the opposite of both.

The sponsored credits expired. Image generation and the reranker stay because nothing else serves them.

It got unwieldy as soon as anything got complex.

I evaluated it once and did not adopt it.

It is billed as API. Not being able to script the agent against itself, or define my own loops and monitors, is most of why the next thing exists.

It owns no state and does no single thing. It reads from every organ, composes them, and decides.

I automate many experiments and filter the best, instead of thinking each one out.

The agent is a commodity. The product is the shell that comes alive when you put one in.

A system that runs, with a configuration space I can search, has done more for me than any single tool.

An agent reporting that its own work is done is not verification.

It gives maximum flexibility with no guardrails except version control.

Sitting at a terminal stopped being a requirement in August 2026.