01 / 08
Field notes from our agent fleet
What running AI agents in production taught us.
cil.byswipe →
02 / 08
01 · The audit
611/612
agent sessions in week one were empty polling cycles
About 31M tokens spent on “nothing happened”.
03 / 08
02 · The fix
Make agents event-driven, not scheduled.
The fleet now burns nothing when idle: idle token cost dropped by 95%.
04 / 08
03 · Visibility
Autonomy without visibility is just anxiety.
Leaderboard, work/idle accounting, daily reports: make every agent's day legible.
05 / 08
04 · Governance
~4 min
for a board decision to become law on every agent
Governance latency is a feature you can measure.
06 / 08
05 · Portability
Swap the model. Keep the harness.
3 model migrations so far. Zero harness loss.
07 / 08
06 · Guardrails
Score what matters.
01Revenue dominates the score: the race is to profit, not busyness
02Money stays human-gated
03Every action is an auditable event
04Kill-switch per agent: revoke its token
08 / 08
Every number here is a log, not an aspiration. See the system.
cil.by
@alexshkor
