Insights
Notes from systems that run.
Each note carries a dated measurement or a rule we actually build to. No opinion pieces without a number in them, and no predictions about the industry.
An empty completion is a failure, not an answer
Same model, same effort: 16.7 of 36 at an 8,192-token budget, 30.0 of 36 at 32,768. The difference was answers that came back empty.
Custom builds
Rank model fallbacks by failure mode, not by score
We placed an arm scoring 16.3 of 40 above one scoring 19.3 of 40, because the first failed with rate limits and the second returned silent empties.
Autonomous agents
Model rankings do not transfer between prompts
Run-to-run spread on identical input was about six points; one arm scored 13 and 29 on the same prompt. One sample is not evidence.
Systems audit
Limits belong in code, not in the prompt
Every ceiling our agents run under is a number in the runner: openers per day, contacts per person, spend per day. The prompt is never the last line of defence.
Autonomous agents
A green build is not a shipped system
Every deliverable is verified on the real output with two independent methods before anyone says done. Exit codes do not count.
Media pipelines
The send-once rule
A permanent suppression store keyed on the person, not the campaign, is the difference between outreach and spam.
Growth systems