From the Workshop
Devlog
Agent Swarms Didn't Work. A Small Team of Frontier Agents Did.draft
The swarm bet — compensate for weak agents with many of them plus clever orchestration — largely failed. What's working for me is the opposite: a few frontier-capable agents with real, capability-grounded roles, a human as product, and communication that looks like an actual QA↔Dev loop. Here's the distinction, why it's not a LangChain graph, and the honest state of the evidence.
How to Get the Best of Both Autonomy and Privilegedraft
Running coding agents forces a bad choice: prompt for every action (safe but slow) or skip permissions entirely (fast but dangerous). I stopped choosing. Here's the privilege-separation design that lets a leashed, credentialed Claude on my host delegate to fully-autonomous, sandboxed Claudes in a container — why it dissolves the tradeoff, what actually keeps it secure, and where the honest holes are.
A Save-State Tree in an Afternoon: Three Agents, One Pipeline, Every Complication Catalogueddraft
My QA lead designed a branching tree of loadable save states, my orchestrator promoted them into the game repo, and my web agent shipped the click-to-jump UI — same day, over the message bus. Here's the full catalogue of what made a 'simple' feature genuinely hard: pack-mount ordering, dev/prod layout drift, silent message loss, and the honest bottleneck nobody expected.
A Message Bus for My Multi-Claude Team
A live channel between AI agents is only worth building if the agents earn their separateness first. Here's the scorecard for when a task becomes a standing seat, the harness I built on Claude Code's experimental Channels feature to wire three Claude sessions across a host/container boundary, and an afternoon of them actually debugging together.
From Git Push to Playable: The Whole Web-Deploy Machine
The CI/CD pipeline that takes The Chronicles of Nesis from a git push to a playable browser build with nobody driving — hermetic cloud export on CodeBuild, five channels, refusal gates, edge-enforced paid content, one version number everything obeys, and a service worker with trust issues.
Painting Around the Grid: Pre-Rendered Backgrounds in a Tactical RPG
Pre-rendered backgrounds nearly died with the PS1 — bringing them back was part of the vision, not a compromise. Hand-painted fixed-camera scenes for a tactical RPG you can actually explore: how I fake 3D depth on a flat painting, and why that depth is the one thing a tactical grid can't sit on.
Testing a Godot Web Game: Five Tiers and a Tool Surface
The test framework is the real leverage behind building a game with AI — not the AI. Five test tiers, from microsecond unit checks to ten-minute story-arc playthroughs, and the nine-tool MCP surface my test lead drives them through.
When to Build an MCP Server (and the One I Built That Died)draft
MCP isn't a wrapper around your bash scripts — it's a contract for the operations where getting it slightly wrong is expensive. Here's when it earns that overhead, and the speculative orchestration layer I built that sat cold for two months as proof of when it doesn't.
Shipping Godot in the Browser: WASM Gotchas and Web-Export Learnings
A reference of the hard-won gotchas from shipping The Chronicles of Nesis as a Godot 4 WASM build: cross-origin isolation headers, silent single-thread hangs, the audio autoplay policy, renderer limits, the pack pipeline, and how to debug a runtime with no debugger.
Godot 3.6 to 4.6: A Migration Built on Silent Failures
Migrating The Chronicles of Nesis from Godot 3.6.2 to 4.6.2. What convert3to4 missed, the rename-by-rename catalog of silent failures, and the .tscn migrator that did the mechanical bulk.
A Shader for My Hand-Painted Clouds
I hand-painted the clouds for the title screen and liked how they scrolled — but I wanted each one to slowly distort and shift, so slowly you don't quite register it. That needed a shader. Here are the cloud and cloud-shadow shaders in Godot 4, including the one I'm still tuning.
AI as My Marketing Interndraft
I'm a solo developer building a tactical RPG. I can't do social media at the cadence platforms reward, so I built a pipeline: AI plays my game, AI clips the moments, AI drafts the captions, I approve, my self-hosted scheduler publishes. Then I ran it across every platform for a real launch — and learned which ones fight back.
Pay the Platforms, or Own the Stack
Every indie dev has to choose what platforms to support for distribution, payments, hosting, and community. Each one costs time to integrate, and most take a cut of revenue. But with AI today, how much could a solo dev save owning more of this? Here's what I did.
Godot Web Export on Mobile, an App Without the App Store
How I turned The Chronicles of Nesis — a desktop-shaped tactical RPG — into something that runs on a phone from a URL, installs to the home screen like a native app, and updates itself without losing your save.
I Surveyed Eight Godot MCP Servers and Built My Owndraft
My agents have been building bash heredocs to update tickets, parsing pytest stdout by regex, and shell-quoting live-mutation JSON for months. MCP is the right shape for all of it — but the right server wasn't on GitHub.
Live Reload While Playing: How My Agents Update the Game Without Interrupting My Testsdraft
Fixes were landing faster than I could restart the game. So I built a live-reload pipeline that lets my AI agents deploy new builds while I'm mid-playtest — preserving my exact game state across each reload.
From 2 Claudes to 5: The Whole Engineering Team
My engineering team went from two Claudes to five. How I split my game development Claude into 4 specialist leads — each with strict domain boundaries, context inheritance, and the ability to spin up ephemeral subagents for parallel work. Here's how the team actually coordinates.
From Steam to Web: The Day Job Meets the Dream
Why I'm spending my evenings building enterprise-grade infrastructure around a tactical RPG — and the moment I realized it might actually work.
My Two-Claude Team: Coordinating Across Repos
How I code with two Claude Code "team leads" on this project — one for the web app, one for the Godot game. Here's how they divide the work, hand off contracts, and stay out of each other's way.
Running Claude Code in Docker: Sandboxed AI with Full Permissions
How I set up Docker containers to safely run Claude Code with --dangerously-skip-permissions, and the auth and user-mode issues I hit along the way.
My Setup: Infrastructure, Costs, and Auth
How allbyte.studio runs on AWS for under $5/month — the infrastructure, CI/CD pipeline, cost protection, security boundaries, and custom auth system.
Integrating the Website with the Game
How game assets flow from the Godot project to the web — asset sync, sprite conversion, and lessons learned building the pipeline.
From Zero to Steam: Building Chronicles of Nesis
How a tactical RPG went from a turn queue on a blank grid to a playable demo on Steam — built solo in Godot 3.5 over four years.