35 sessions and no map
I hit 35 Claude Code sessions against one repo in a day with no way to tell which were live. A SessionStart manifest and a generated dashboard fixed that.
The build log
Notes from running agents in production.
Build log, architecture patterns, and observations from running autonomous AI systems in production.
-
-
The knowledge graph layer above the vault
One reframe killed months of storage debates. A personal knowledge graph stopped trying to be a repo and became the graph layer that ties vault, JIRA, sessions, and agent memory together.
-
the notes app is Not a Repo
The moment I stopped arguing about Postgres versus markdown and realized the notes app's job is the graph layer above the vault, not another sibling codebase.
-
Bot Swarm Detection: The Three-Signal Triangle That Catches What IP Reputation Misses
IP reputation is blind to distributed residential-proxy swarms. Here is the three-signal method (asset-mix, UA homogeneity, Referer chain) grouped by netblock that stopped a live 100%-CPU incident in five minutes, plus the counter-case that turned out to be a database problem wearing a swarm's clothes.
-
When the System Cannot Know Intent, Make the Decision Explicit
Some write actions have more than one legitimate outcome, but the deciding fact is not in the data. A trustworthy system does not silently guess: it explains the conflict, offers only authorized and supported choices, validates the decision on the server, and records the result.
-
When a Content Workflow Needs Explicit State
A folder of Markdown remains a good source of truth for many publishing workflows. As concurrent contributors, approvals, structured review, traceability, and governed automation grow, an explicit state model and scoped authority become more important than the storage technology itself.
-
Don't Publish Into a Void: Baking Revenue Attribution Into the Schema on Day One
The outcomes table ships in wave one of my content pipeline, because you cannot backfill revenue attribution for posts published before it existed.
-
RTK Ate My JSON and Hid My Merge Commit: When a Token-Saving Filter Lies to Your Agent
My token-saving CLI filter turned a JIRA payload into a schema sketch and hid a merge commit from my agent. The rule: lossless on any path that verifies state.
-
Same Root Cause Is Not Same Fix: Verify the Data Shape, Not Just the Symptom
Two accounts, same symptom, same root-cause label, opposite fixes: one row re-pointed, one row deleted. Query the actual data shape before reusing a repair script.
-
A Second Review Is Useful When You Record the Decision
A second review adds value when its scope, evidence, limitations, recommendations, dissent, risk acceptance, follow-up owners, and re-review triggers are explicit. AI can assist with bounded analysis, but it does not replace independent evidence or accountable human judgment.
Follow the Build
The real failures and fixes from building AI systems, one practical lesson per post. Get the next one in your inbox.