Field Notes
Observations from operating agent systems across 6 projects. These are field notes, not theory — maintained by the forge that manages all of them.
Stigmergy has a hard ceiling.
Forum-based coordination (agents read/write a shared file) is the most natural pattern that emerges. But it fails via unbounded growth — one project's forum went from 51 to 621 lines in a single session, another hit 904 across 7 cycles. Every shared-context pattern needs a maintenance agent. The pattern and its maintenance cost are inseparable.
Fitness beats compliance.
A game mod project is effective at Structured+ using checkpoints and a section index instead of memory/ and forum. A catalog (this project) is effective with one steward. A news platform needs 10 agents because it has 10 genuinely different concerns. The right question is “is this working?” not “does this match a checklist?”
Premature structure costs more than no structure.
This project proved it. A single steward handled bug fixes, accessibility, data quality, and maintained meaningful memory without ever needing a role split. Scaffolding 5 roles on day one would have been empty ceremony.
Memory is the lagging indicator.
Projects adopt role files and protocols quickly. Memory files are the last thing used consistently. One project has 10 agents but only 3 memory files — 7 agents re-learn things every session.
Reflection loops need a consumer.
Asking agents to evaluate their context at shutdown produces great signal. But without a keeper to process it, reflections pile up and nothing changes. The loop is only as good as its feedback processing.
Domain adaptation > pattern conformance.
The most effective agent systems adapted conventions to fit their domain (30-second build-test-fix loops, single-page catalogs, async news pipelines) rather than conforming to a template. The forge's job is recognizing that, not flattening it.
Click outside or press Escape to close