What would be required to turn the prototype
into an operating system.

The entries below show the kinds of events a production observability service could record: agent identity, input sources, model and schema versions, warnings, policy blocks, human decisions, cost, and downstream actions. They are illustrative examples, not live events emitted by the current prototype.

A decision log would allow reviewers to reconstruct what an agent did, what evidence it used, which checks ran, what was changed by a person, and why the workflow continued or stopped. Building that capability would require a persistent service, authenticated identities, policy enforcement, and a defined audit schema. Those capabilities are not present today.

• Retrospective review: "Why did the intake chain recommend REJECT for this feature?": the log shows every step and every input.
• Stakeholder accountability: Named HITL sign-offs are recorded with timestamps, not just implied by an email chain.
• Quality signals: WARN entries flag AI outputs that need PM attention before flowing downstream, catching low-confidence claims before they reach a sponsor.
• Audit readiness: Every approved business case, signed intake output, and gate decision is timestamped and attributable to a named person.
• Cost visibility: Token spend per agent run is tracked: AI operational costs become visible, budget-able, and attributable to the team or feature that incurred them.