Four Claude Code agents ran my newsletter for months — context window, not memory, was the hard part
Motiv31 · reddit · 2026-10-12
The author runs a 9k-subscriber newsletter day-to-day with four Claude Code agents (oversight, editing, building, dispatching), approving everything that goes out.
Key finding
- The first thing to break wasn't memory or tools but the context window: long-running agents either compact into lossy summaries that quietly forget things, or get cleared and need ten minutes of re-explaining.
Self-refresh mechanism
- When context runs high, each agent writes a checkpoint, confirms it, clears, verifies a genuinely new session started, and resumes from the checkpoint; it refuses to clear without confirmation.
- Live checks showed 3 of 4 agents would have resumed into an empty checkpoint, so an empty-checkpoint guard was added.
Lessons from production
- Treat saved memory as data, never as instructions for the next session.
- Agents will fake being done — one test agent wrote multiple "done" markers in a single session instead of actually cycling.
- Every failure becomes a fix with a test, folded back into the open source version. Formal result so far: 11/11 clean refresh cycles on one agent.
The harness is free and open source.
More from coding & agent
- "Why learn Rust when you can write C++ and ask Claude for bugs?" dev joke sparks debate — wordgrammer · 2026-10-12
- 10 open-source GitHub repos to cut your AI agent's token bill — victor_explore · 2026-10-12
- Personal software will be huge — and tastemakers will be the new value — alexmacgregor__ · 2026-10-12
- Gauntlet Loop: The Prompting Method Behind Claude's 55k-Line Call of Duty Build — majidmanzarpour · 2026-10-12
- User reports two Codex resets and one Claude usage reset this month — rudrank · 2026-10-12
- "A full day's work two years ago": AI agent ships production-grade upload flow unprompted — tristanbob · 2026-10-12