What survived production.
In July I wrote about a hierarchy of coding models, automatic memory retrieval and agents choosing their own compaction points. By October, I had retired substantial parts of all three. The useful parts were smaller: a clear task, evidence attached to a claim, and a record that lets the next session continue the work.
The task contract stayed
Conductor economics described a frontier model planning work for cheaper executors. I retired that default hierarchy in September. Model choice now follows the task and the active session. The July pricing arithmetic describes that earlier setup, not a current recommendation or a measured saving for today’s workflow.
I kept the written brief: the problem, relevant decisions, acceptance criteria and files an agent may change. Parallel writers need separate worktrees and clear ownership. Their output still needs review against the actual diff. None of those requirements depends on one model being assigned the permanent role of planner or executor.
Memory became something to ask for
The July memory design injected retrieved notes at the start of work. The current system serves 21 projects through on-demand queries: an unfamiliar failure, tool or API is a reason to look. A retrieved note supplies a lead and its source; the agent checks the source before treating the note as true.
The knowledge vault is now a plain Markdown folder, read directly. Its Obsidian service and mirror were retired in October. The freshness job is report-only: it flags potentially stale, superseded or duplicate memories without changing or deleting them. I kept retrieval and provenance, while removing the assumption that every session benefits from automatic injection.
The checkpoint outlived its trigger
Self-compaction let an agent write a ledger and choose when to compact. I retired the system on September 1. Its context-ceiling estimate drifted on the million-token window, causing sessions to compact at about half their available context.
A useful checkpoint still records the active task, decisions, exact file paths, checks already completed and the next unresolved step. That format remains useful when the harness controls compaction. The old hooks and their 215 regression checks stay public as an inspectable implementation; passing those checks did not make the context estimate reliable forever.
What I now ask of an automated decision
I want to know what evidence supports it, what code limits it, and how I would notice a bad result. In test selection, the replay exposes the extra coverage bought by model questions, and an audit runs skipped files after merge. In bot mitigation, a confident wrong answer on organic browser activity is visible alongside the deterministic veto that keeps it out of enforcement.
These are different systems, but both put an inspectable decision around the model. The July essays remain here with their original dates and current-status notes. Keeping the history makes it possible to see which ideas earned continued use and which assumptions stopped holding.