Six engineering rules for reliable agent systems.
Back to OverviewEvery write action is followed by a read or verify step to confirm success.
Capabilities are taught progressively: foundation, conditional logic, chaining, natural keys, async, synthesis.
Tools are atomic, single-purpose, and deterministic. State belongs to tools, not the agent.
Reports are generated from structured templates, not freeform text.
Golden files, unit tests, and validation suites prove correctness before release.
After natural-key capability, a human reviewer issues binding GO/NO-GO with audit trail before progression.
These principles emerged from high-stakes deployments where deterministic, auditable behavior mattered more than model size.
HITL is intentionally a late-stage gate: meaningful review requires outputs that humans can ground-truth against source data.