AlgoSu Tech Blog
The build & ops log of an AI-agent algorithm-study platform built by a solo developer. From 12-agent orchestration, k3s ops, CI/CD, and ADRs to production debugging — the problems and decisions behind a service with real users.
AlgoSu at a glance
20+
Active Users
A service with real users in production
12
AI Agents
Multi-agent orchestration
6
Microservices
Service-boundary architecture
23
CI/CD Jobs
Validation pipeline for AI-generated code
202
ADRs
The SSOT of decisions & ops records
∞
Zero-downtime
Fast iteration without downtime
New here? Start with these
Grasp the core of AlgoSu through a few signature posts.
- How to Control 12 AI AgentsHow a 12-agent orchestration is structured
- Safety Net for AI Code — Building 15 CI/CD Jobs in PracticeThe strategy for shipping AI-generated code to production
- Memory Fades, Documents Pile Up — A Sliding Window for Agent ContextTaming agent-context growth with a sliding window
- Toward a Model-Agnostic Harness — Adopting a CriticBreaking model lock-in with cross-AI review (Critic)
Architecture Decision Records (ADR)
The single SSOT of decisions, implementations, and verifications cited by every post — 202 ADRs visualized
A decision log of problem → options → decision → verification
Recent Posts
- Retrospective
The Art of Deletion
Deleting a feature is scarier than adding one. You never know what might follow. Building and removing the ADR graph and search in AlgoSu taught me that safe deletion is a craft with its own method.
retrospectivedeletionyagnidead-coderefactoring - AI Agent
Which Model, When, Where — A Strategy for Targeted Premium Deployment
I switched to Fable 5 the day it launched. Six sprints later, what I'd gained wasn't performance — it was strategy. A record of when to reach for a premium model, and why that judgment is the same thing as paying down technical debt.
retrospectivemulti-modelmodel-selectionrefactoringcost - Platform
I Fixed an Outage That Never Happened — When a Static Manifest Judges Runtime
I was certain alerts had been silently dropping for six days. I diagnosed it, fixed it, even merged it. But the file I'd been reading wasn't even deployed — it was a mirror. A six-sprint record of letting a static manifest pass judgment on runtime, and being wrong.
monitoringobservabilitygitopsincidentdebugging - Retrospective
I Chose to Spend More Tokens — Building an ADR View for Humans
What I learned by separating Markdown ADRs for agents from HTML ADRs for humans.
agentic-systemadrcontext-engineeringdocumentationllm - AI Agent
Memory Fades, Documents Pile Up — A Sliding Window for Agent Context
Across 152 sprints, the ADRs, MEMORY, and CLAUDE.md all swelled together. Yet handing the entire pile to an agent every session was like pouring more water into a full bucket. A record of trimming agent context with a single rule borrowed from the sliding window algorithm.
agentdocumentationsliding-windowcontext-optimizationclaude-code - AI Agent
Toward a Model-Agnostic Harness — Adopting a Critic
I brought in Codex — as an addition, not a replacement. After 17 rounds of review, I realized what I needed wasn't a different model. It was a system freed from dependency itself.
challengemulti-modelcodexcriticharness - CI/CD
AlgoSu CI Refactoring
Starting from a Channel.io blog post — auto-merge, composite actions, coverage gates, and pre-consultation across five sprints. A record of how one reference article turned into a real pipeline overhaul, and why two sprints that shipped zero lines of code were still wins.
ci-cdgithub-actionsai-devrefactoring - Platform
Baekjoon Is Gone?
On the day I saw Baekjoon's shutdown notice, migrating AlgoSu's problem source to Programmers taught me something I didn't expect — not a new technology, but an attitude about building services that lean on external platforms.
retrospectiveexternal-apiplatform-migration - Backend
Why Does My Session Keep Dropping? — Debugging 4 Layers of Desynchronization
JWT TTL is 2 hours, but users get logged out after 1 hour. Cookie maxAge, sliding refresh, frontend timer — 4 layers each had their own clock. Here's how we unified them into a single source of truth.
sessionjwtcookiedebuggingnestjschallenge - Retrospective
Looking Back on 67 Sprints
A project that started from curiosity about multi-agent systems became a full platform over 67 sprints. An honest retrospective of good decisions, deferred ones, and the journey from coder to builder.
retrospectivesolo-devgrowth - CI/CD
Safety Net for AI Code — Building 15 CI/CD Jobs in Practice
The faster AI wrote code, the tighter the verification had to be. A hands-on account of building 15 CI jobs — from secret scanning to GitOps locks.
ai-devci-cdgithub-actionsgitops - AI Agent
How to Control 12 AI Agents
As AI agents multiplied, control became the problem. An arbiter, a 3-echelon hierarchy, SSoT rules — a firsthand account of building a governance structure solo.
ai-devagentgovernance - Architecture
MSA Design: Humans Decide, AI Executes
A hands-on record of MSA design where humans make architectural decisions and AI carries them out
ai-devarchitecturemsakubernetes - AI Agent
AI Agent Orchestration in Practice
A hands-on account of building an MSA solo while growing 12 AI agents into teammates
ai-devagentorchestrationclaude-code