LOADING
The difference between a demo and a system is everything below. Most of it is invisible when it works, which is exactly why it gets skipped — and why so many agent deployments quietly fail in their third month rather than their first week.
So here they are, made visible the only way that works: switch one off and watch what it was holding back.
Every side-effectful action carries a key. Replaying it is a no-op.
Every agent action writes one row to an append-only event log. Monitoring, client outcome metrics, regression baselines and true project margin are all derived from that one table — not from four separate pipelines that disagree with each other.
Most companies build monitoring first and discover eighteen months later that they cannot prove client value, because the events were never designed for it. We built the schema before we built this website.
A linear chain serialises work that has no dependency on itself. A graph fans out and joins. But parallelism applied everywhere is the fastest route to corrupted state, so the split is deliberate rather than automatic — and every join carries a reducer.
Never raised without a reason written down. A threshold that drifts upward quietly is a circuit breaker that will not trip when it matters.
| Setting | Value |
|---|---|
| Max parallel branches per node | 8 |
| Per-provider concurrency | Published rate limit, minus 20% headroom |
| Retry policy | 3 attempts, exponential backoff, then escalate |
| Action cap per tenant | Per hour, at 3× observed peak from the Baseline Week |
| Spend cap per tenant | Per day, hard stop with an alert at 70% |
| Eval regression tolerance | Zero on safety and escalation cases |