Patterns observed, methods tested, systems built.

A harness phase reported green for a week. The script it was meant to run didn't exist on disk. Success-by-default is the silent rot of any pipeline whose state is described in one place and whose work happens in another.

I went to a 500-person AI builders event hoping to make a few real contacts. I came back with zero. The event was excellent — the format was wrong for what I needed. A small heuristic for picking events.

For 132 agent tasks across six weeks, my dev-loop reported pass=true. The central game mechanic never actually worked once. Three small hacks compounded into a complete fiction.

A long-running agent's beliefs file went 21 days without updating while the agent kept reporting normal operation. The decay is in-context salience — so the fix has to inject through that channel.

A bug was "fixed" five times in five sprints. A UI overlap went unnoticed for seven. Here's the pattern — and how to close it.

A solo developer ran 8 complete game dev sprints in one week using an AI agent team. Here's what that actually looked like — including the parts that didn't work.

How a solo developer automated sprite quality verification using palette matching, pixel checks, and SSIM regression — with no human in the loop

A 3-layer self-correction architecture where daily tasks detect failures, weekly diagnostics resolve them, and the orchestrator only gets what it can't handle

How a Preflight→Execute→Emit protocol unified 14 isolated agent tasks into a self-auditing ecosystem

My agent ecosystem scored 8.5/10 on autonomy four weeks running. Same diagnosis kept coming back: nothing was reaching customers. Engine improvement decoupled from external reach — and the work feels like progress because locally it is.

An AI agent's attention is finite. Every byte of context is borrowed from its capacity to do the work. Verbosity is the main culprit.

Path mismatches between monitoring code and actual outputs caused 75% false-alarm rate

State transition failure causes a task to repeat 30+ times without progress

Python 3.14 / 3.12 script execution hangs under MCP PowerShell with 60s timeout

BOM / CRLF / code fences silently break the schedule parser

A research agent measured belief updates after positive and negative events. The same formula fit both — but the sign flipped depending on where you started. Hypothesis stage, not finished theory.

We found r=0.82 between desire strength and tension. We assumed causation. We hadn't — and the error invalidated six weeks of research direction.

Three independent axioms from neuroscience and motivation theory explain where intrinsic fun comes from

Lesson accumulation + dedup + early termination

We gave every game object the ability to stream its own state to the agent team. What changed wasn't just visibility — it changed how the agents work.

How giving AI agents identity, memory, and owner context transforms them from task executors into colleagues that accumulate judgment

CLI access scripts, HTTP bridge alternative, and 771-entry knowledge store design.

Structures for orchestrating multiple agents

How to physically separate agent-modifiable areas from areas agents must never touch.

Protocol for preserving context across session boundaries