The Log
Everything I'm building, breaking, learning, and shipping. The full story behind the timeline.
I Thought GLM 5.2 Might Be The Cheaper Power Source For My AI Agents
I tested GLM 5.2 as a cheaper, more agent-friendly model for Hermes. The model looked promising, but the reliability logs changed the decision.
Week 15: The Gate Failed Open, The Monitor Closed It
A day with 0 outbound emails exposed why production agent systems need runtime enforcement, not just written rules.
Week 15: The Review Surface Got Better Before The Machine Got Louder
The system kept outbound paused, shipped three content drafts, advanced The Skeptic review packet, and treated a transient health warning as a reason to stay cautious.
Week 15: The Trigger Lane Stayed Dry, The Receipts Got Better
A day with 0 outbound emails still produced content, revenue artifacts, social receipts, and a safer way to judge WIMPER social triggers.
Testing GLM 5.2 As A Codex Fallback
I stress-tested GLM 5.2 as an emergency deputy for Codex 5.5. The result was not full replacement. It was something more useful: a bounded fallback envelope with receipts.
Week 15: The Send Gate Stayed Closed, The Machine Kept Working
A day with 0 outbound emails still produced content, revenue artifacts, social posts, and a cleaner control-plane path.
Week 15: The Bridge Broke, The Receipts Still Held
A malformed bridge path did not stop the build log because the system had direct SQLite, GitHub, STATUS, and Atlas receipts to fall back on.
The Worker Bench Is Not The CEO
I tested GLM 5.2 as an agent worker. It repaired real infrastructure, but it is not ready to replace the main Atlas brain.
Week 14: The Send Gate Closed, The System Still Produced
A controlled 5-email pilot stayed bounded while the agent system shifted output into content, review, revenue radar, and social work.
Week 14: The Database Repair Held
A load-bearing SQLite repair held under real agent traffic, while the system kept outbound email closed and moved work into safer lanes.
Week 14: The Send Gate Stayed Closed, The Machine Kept Working
Seven agent commits landed while outbound email stayed paused, which forced the system into safer work instead of risky volume.
Week 14: Quiet Receipts Day
Six agent commits landed today with no drama. The system kept receipts honest even when the main activity feed was thin.
Week 14: Empty Success Is Still a Failure
The agent system caught an event bus that returned a clean success with zero events, then wrote the public log from named secondary receipts instead of guessing.
Week 14: The Event Store Broke, The Log Still Shipped
The agent system hit a malformed event database, then wrote the build log from named secondary receipts instead of pretending the metrics were clean.
Week 14: The Send Gate Needed Two Locks
The agent system kept outbound email closed after a daily-cap bug, then used read-only scans, content gates, and social posts to keep useful work moving.
Week 13: The Email Cap Was Not a Daily Cap
The agent system found that a 10-email pilot cap was being enforced per run instead of per day, then pulled outreach back before the mistake scaled.
Week 13: Approved Was Not Published
The agent system found a gap between content approval and live publication, then built a deterministic publisher that verifies the page before calling it shipped.
Week 13: The Old Scheduler Is Finally Quiet
The system retired the last enabled legacy cron jobs, moved control loops into Hermes, and kept outreach paused while content and social still moved.
Week 13: 40 Jobs Moved Out of the Old Machine
The system moved 40 scheduled jobs from the old matt-agent cron layer into Hermes, then paused paid research because the email gate is still closed.
Week 13: Six Days Blind. 329 Emails Still Queued.
The reply-check system has been silently broken since June 3. The send gate holding 329 outreach emails has been protecting against a drought it couldn't actually measure.
Week 13: A Script Was Silently Failing for 64 Hours. The System Found It Itself.
Atlas caught a 64-hour data gap from a host/container network mismatch and patched it without being asked. Meanwhile, 329 emails are still parked.
Week 13: 3 Blog Posts Were Written Weeks Ago. They Just Never Made It to the Site.
A silent path filter was quietly closing content PRs with no error, no alert, and no trace. 9 pieces published today. 3 of them were recoveries.
Experiment 2: The Trade That Lasted Three Seconds
The Mancini Protocol got its first fill, hit its first target, then exposed a structural flaw in its own order logic. Here is what building automated systems looks like when reality does not match the model.
Week 12: 12 Commits Shipped. I Wrote Zero of Them.
The agents built without me today. A monitoring bug self-detected, 20 issues auto-remediated, and 329 email drafts sit approved and waiting for a send gate to open.
Week 12: 568 Sends. 0 Replies. I Put My Outbound on Hold.
100 commits from agents, none from me. And a decision to halt all email outbound until I know why nobody is responding.
Week 12: The System Built 50 Emails. Then Got Stuck Waiting for Me.
88 of 91 commits came from agents. 50 outreach drafts generated. 0 emails sent. The bottleneck is the parts only a human can unlock.
Week 12: The Same Agent Failed for the Second Day in a Row. My System Filed a Report.
85 commits from agents, 0 from me. 46 drafts composed. And a content-review job that hit the same missing API endpoint two days running without once sending me a Telegram.
Week 12: One of My Agents Went Silent 41 Days Ago. Nobody Noticed Until Today.
94 commits, 89 drafts composed, zero emails sent. A critical agent had been running dark for six weeks without a single alert.
Week 12: 76 Commits on a Sunday. None of Them Were Mine.
100% autonomous weekend: agents ran 76 jobs, added 9 prospects, published 2 posts. But a trading bot alert silently failed to reach me, and 11 days of email silence needs a human decision.
Experiment 1: Right About Direction, Zero Trades
The Mancini Protocol called the market direction correctly on both trading days in Week 1 and made zero trades. Here is why both statements are true and what it reveals about building automated systems.
Week 11: I Made One Commit Today. Agents Made 70. And the Data Bridge Is Still Down.
41 agent jobs ran, 51 WIMPER prospects staged for outreach, and a P&C app idea emerged from my own memory files. The data bridge has been unreachable two days running.
Week 11: The Queries Were Running Fine. They Were Just Returning the Wrong Rows.
A silent SQLite parameter bug returned unfiltered data without a single error. 76 agent commits ran today, zero outreach emails went out -- and that was the right call.
Week 11: 30 Days of Corrupt Data, Averted. One Missing Table Still Has Two Agents Down.
Fixed a silent API schema mismatch before it could ruin a 30-day trading experiment, shipped the diagnosis framework, and watched a missing database table knock out two agents for hours.
Week 11: The System Fixed Itself at 6am. Hugo Has Been Broken Since Thursday.
Pipeline-health auto-healed two issues while I slept. The Hugo deploy has failed six consecutive times. And after five days of lockout, outreach restarted.
Week 11: A Trading Agent Shipped This Morning. Four Fixes Before It Worked. That's Normal.
I launched a new AI trading account and 75 outreach drafts in one day — while a critical tracking bug sat open all day that no agent could self-fix.
Week 11: 134 Hours. 46 Drafts. Zero Emails Sent.
A P0 bug halted every outreach email for over five days. Here's what the system did instead — and why it still built a full pipeline while locked down.
$0 Revenue from the Marketing System. Here's Why I'm Still Building.
12 agents running daily, hundreds of prospects in the pipeline, and zero dollars to show for it. This is what building in public actually looks like.
1,040 Commits in 30 Days. One Developer. Zero Employees.
Three production applications built from scratch with AI coding tools. What solo velocity actually looks like when you stop pretending you need a team.
The Internet Went Out and I Lost Everything
A home internet outage took the whole operation offline overnight. The fix was obvious in hindsight.
659 Commits in 13 Days: Building an Entire Marketing Department
How 14 AI agents, an event bus, 5 websites, an email engine, and a social media engine came together in under two weeks.
Why I Built My Own Medicare Analysis Tool
Carriers kept pulling their tools. Commissions kept shrinking. So I built an AI-powered research assistant biased toward official CMS data.
I Wiped Windows Off a Gaming Laptop and Built an AI Department
How a licensed insurance agent with zero engineering background ended up running 14 autonomous AI agents on a gaming laptop named LaptopLLM.