Experiments
+Controlled AI workflow observations with setup, evidence and limitations.
-
+
++ +
We Removed Middleware From Our WordPress AI Workflow — What the Native REST API Already Gave Us
A practical case study: we moved routine AI-driven WordPress work onto the native REST API, tested posts, pages and media end to end, and kept privileged plugin deployment separate. Read the article →
-
+
++ +
We Ran 12 GPT-6 Luna Low Sessions at Once — Here’s What Happened to CPU and RAM
One owner-reported, single-machine snapshot: CPU ~39–71%, RAM ~25.6–26.0 of 31.8 GB, GPU ~10%. This is not a concurrency ceiling or controlled benchmark. Read the article →
-
+
++ +
Building AI Tools We Needed for Real Workflow Questions
Three focused browser tools address recurring decisions about agent rules, context policies and measured AI costs; adoption and savings remain unmeasured. Read the article →
-
+
++ +
AI Workflow Acceptance: A Heartbeat Is Not a Verified Result
Separate worker liveness, provider activity, durable progress and verifier acceptance; no production throughput rate is claimed. Read the article →
-
+
++ +
Measure the Final Request Boundary, Not the Visible Prompt
A reproducible method for capturing assembled messages, tool schemas, retries and provider-reported usage without attributing tokens by guesswork. Read the article →
-
+
++ +
WAIT_AUTH: Separate Authentication Recovery from Resource Waiting
Treat credentials and provider capacity as separate wait states, preserve the checkpoint and measure human recovery rather than guessing. Read the article →
-
+
++ +
A Plan Change Coincided with OpenCode Auth Failures: What the Logs Actually Show
Redacted logs show four auth-invalidated records across two sessions; timing is measured, but the plan-change cause and resumed jobs remain unknown. Read the article →
-
+
++ +
WAIT_RESOURCE: What a Provider-Free Resume Test Can—and Cannot—Show
A fake-adapter wait test can exercise checkpoint and retry invariants; it cannot establish live provider recovery or accepted output. Read the article →
-
+
++ +
How to Detect a Stalled AI Agent Without Killing a Legitimately Long Job
Separate process liveness, provider activity, semantic progress and acceptance so quiet legitimate work is not mistaken for a stall. Read the article →
-
+
++ +
A live heartbeat is not model progress: lessons from a stalled canary
A reported >60-minute canary stall shows why wrapper liveness, provider activity, task progress and verification must be tracked separately. Read the article →