Observe + prepare
Agents may read approved sources, analyze, summarize, and create drafts that remain inside the workspace.
- Research and synthesis
- Internal forecasts
- Draft documents
GovernanceFast where safe.
Fast where safe.
Gated where final.
04 / GOVERNANCEBOUNDARIESAPPROVALPROVENANCEESCALATION
Three autonomy levelsAuthority increases
Authority increases
only with evidence.
A workflow should default to the lowest authority needed. The action type—not confidence theater—determines whether a human gate is required.
Reversible execution
Agents may perform bounded actions that can be easily inspected and rolled back inside an approved workflow.
- Tag or route records
- Create internal tasks
- Refresh approved models
Human-gated action
One-way doors stay locked until an accountable person sees the evidence and explicitly approves.
- Send or publish
- Purchase or commit funds
- Delete or change customer data
Decision receipt
Every consequential action leaves a readable trail.
A receipt should make the decision reproducible without exposing private chain-of-thought. It records inputs, evidence, checks, authority, and outcome.
See it inside a workflow ↗RECEIPT / HG-2026-041
Approve dormant pipeline campaign
- Objective
- Reactivation
- Evidence
- CRM + policy / 5 sources
- Prepared by
- Growth + Revenue
- Risk class
- External contact
- Policy checks
- Passed / 7 of 7
- Approved by
- Human owner
Fail fast, fail loudUncertainty should surface.
Uncertainty should surface.
Not disappear.
The system pauses when evidence is missing, sources disagree, policy fails, or authority is unclear.
CONDITION
SYSTEM RESPONSE
HUMAN SEES
Missing evidence
Stop the dependent step
Exact missing source
Conflicting sources
Preserve both; do not guess
Conflict and freshness
Policy violation
Lock the action
Rule and remediation path
Tool failure
Return an explicit error
Failed step and safe retry
Authority unclear
Escalate to accountable owner
Why approval is required
Design control before automation.
Start by marking the one-way doors, then decide what the system may safely prepare and execute around them.