Part 5 · 1 chapters · ~8 min

Guardrails

A written policy of autonomous, assisted-only and never, the money-flow line around ledger, holds, fees, limits and providers, production actions kept human by credentials, data boundaries for prompts, and cost and quality harnesses that decide how far trust extends.

6

What stays human, the money-flow line, cost and eval harnesses

explicit, enforced, measured
  1. A policy that sorts work into autonomous, assisted only, and never.
  2. The money-flow line: assisted mode, two reviewers, tests approved by a human.
  3. Production actions stay human, enforced by credentials.
  4. Data boundaries follow data classification.
  5. Cost is measured and budgeted.
  6. Quality outcomes decide how much trust to extend.
code
# .github/CODEOWNERS + a CI check: the money-flow line, enforced
/src/ledger/          @acme/ledger-owners @acme/payments-leads
/src/payments/holds/  @acme/payments-leads @acme/risk
/src/pricing/         @acme/pricing-owners @acme/finance-eng

# ci/money-line.yml: PRs touching these paths need 2 human approvals and a human-authored test change
on: pull_request
jobs:
  money-line:
    if: contains(join(github.event.pull_request.labels.*.name), 'agent-authored')
    steps:
      - run: scripts/check-money-paths.sh   # fails if money paths changed without 2 human approvals
                                             # or if test files under those paths were only agent-edited
instructions are not controls
"Do not modify the ledger" in a prompt is a request. A read-only credential, a CODEOWNERS rule and a CI check are controls. Put the important lines into the second category.
GUARDRAILS
what stays under human control, where the money-flow line sits, and the harnesses that keep cost and quality measured
swipe the figure sideways, or tap expand for full screen
1/6
the policy
The policy: a short document (ai-policy.md) that says which work agents may do autonomously, which only in assisted mode, and which not at all; who reviews agent-authored changes; what data may be sent to which model providers. Written once, referenced from the standing context, reviewed quarterly.